DS isn't better than Sol and Gemini in the benchmarks, but after working with them all for a while now I'm beginning to understand why people find DS _feels_ so much better.
First, it's the tok/s and TTFT from providers. Deepseek blazes through work - at quite good quality - and you don't find yourself waiting for your agents to finish.
Second is the openness. Being able to see the CoT and not being randomly blocked for cyber review. You're never wondering if it's stuck or if it will randomly fail. Combined with the great cost value.. My next subscription will probably be Z.ai
Feels better to me in the sense that it doesn't try to be too clever. When I ask it to do things it actually generates code that a developer would write. That still lets me build software in layers like I do mentally. And it keeps you in control of where abstractions are.
Some people would say that you don't need this at all with the frontier llms. But when extending existing software or writing software that fits into an existing ecosystem, I find I still need to offer a lot of guidance. And it's better that the context window isn't flooded with code that was reverted after I looked at it.
> Some people would say that you don't need this at all with the frontier LLMs
Hit the nail on the head. Development methodologies pivoting at the whim of frontier models have been one of the things that have been worrying me as of late.
it doesn't worry me too much tbh. This is really just cutting corners and going this route leaves you with some kind of additive ship of theseus piece of software you don't understand and require your frontier model to work with, which just... becomes more additive and gross.
right now we're in some proverbial heyday of everyone seeing how absolutely grand of a thing they can make with frontier models.
over time i expect some old design philosophies to take hold with people who use LLM's successfully and find a place again paired with developers who understand why such layered methodologies are important... because this whole AI thing wasn't the silver bullet CEO's and crazy execs thought it was. They will eventually get sick of the support burden of badly designed stuff. i anticipate a bit of pendulum swinging back towards "still using AI, but we need smart people to do it because it's expensive and we want it done right". Right now a lot of companies are being extremely reckless imo and they haven't fully realized/felt the cost implications of trainwreck software yet.
> haven't fully realized/felt the cost implications of trainwreck software yet
As long as the API is private or not dependent on by automated systems and you have good test coverage, you can get away with bad code IMO since you can always refactor. If it looks stupid but it works (and can be refactored/fixed/improved later if needed) it aint stupid.
IMO the fundamental principles and "conflicts" are still the same, but many have forgotten about them with all the hype and sense of speed. The industry will slowly rediscover them.
I just use a suite of tests to enforce the mechanical standards (dry, encapsulation, etc) and arch documents for shape + reviews of new work against the two. I’m not seeing any significant drift.
Yea, this is one thing that gets me. Vibe coding and freestyle prompting gets wildly different results. Starting from a set of specs and expectations routinely gets me reviewable and we'll tested code. Sure, running the full test suite alot slows things down and uses more tokens, but when I hit a five hour or weekly window, I just roll up my sleeves and start working without the model because the code is layed out the way I expected and is designed to fit my meatbag workflow and pace.
> Feels better to me in the sense that it doesn't try to be too clever. When I ask it to do things it actually generates code that a developer would write. That still lets me build software in layers like I do mentally. And it keeps you in control of where abstractions are.
yes!So it's not just me. In terms of how intelligent they want to be OpenAI is at the top, then Claude. And actually when I was playing with Sol it meant that many tasks it would just try to do something else because it thought I don't know what I want. Even trying to nudge it back in the right direction wouldn't work.
I guess the models just have the cultures of their creatures.
Interesting. I’ve been using GLM 5.3 Flash for a couple months and found it to be great. What has it been doing that you found to be “too clever”? It sometimes gets stuck in loops. I once asked it to do a code review of a simple pattern and it went off on a 45 odyssey trying to analyze too much. But other than those, it’s been a pretty good daily workhorse. And you can’t beat the price. Even Deepseek 4.1 Flash is more expensive. And I switched to GLM 5.3 Flash from DS 4 Flash, my previous daily driver, which I also found quite good. Anyway, my point isn’t to start a “which model is better” subthread, but was just curious what behavior you’ve seen where it’s “too clever.”
Anecdotal but I was doing regular adversarial reviews of new work using about 5 models (mostly the open Chinese ones like DS, GLM, Qwen, Kimi, etc) and GLM routinely took much longer for similar results and the same instructions.
It would just loop on things and go down rabbit holes more. DS flash has been so consistently good I've kinda settled on it for 90% of my work.
I still use adversarial review for bigger changes where I want to make sure any issues are caught early, but I genuinely feel as tho if no new models ever get made I'd be fine with just using DS forever.
It wastes my time. A lot! In different ways. That's the tl;dr. Really.
Keeps repeating things I settled as accepted or rejected. But what really gets on my nerves? The language it uses – that aweful, useless typical LLM languages! No AGENTS.md instruction, no per session instruction, no just two-prompts-prior instruction seems to work, hell sometimes instruction regarding that in same prompt doesn't work. GOOD/BAD example along with instruction don't work. That wastes time and zones you out! Not to mention the fatigue of to and fro. DS has been working much better. Hell Gemini 3.1 Pro works so much better than GLM. I don't have extended experience with OpenAI and Anthromic models but the free claide.ai works quite well and better than GLM 5.3 (not even flash). Albeit I must say I rarely use claude.ai for coding, but technical tasks? All the time!
GLM's best response: "No excuse. My fault. You exactly asked me not to do exactly that and so did the AGENTS.md. I failed". Love it. Except that if I press I will get that response every 5 prompts.
I briefly tried GLM 5.3 Flash and the first thing I noticed was how slow it was followed quickly by its single-mindedness. With DS 4.1 Flash I find that it understands that not all the detail in my prompt is important. Sometimes I like to give a quick example of what I'm looking for and it can take the hint whereas GLM will seemingly reshape its worldview around this tiny detail.
DS just handles ambiguity and my bursts of detail with much more grace. I'm only using "max effort" and the CoT is reasonable compared to GLM.
Z.ai does offer access to DS AFAIK, so I guess that's what they mean? Maybe the feeling is that it provides more value that DS's own platform since you also get GLM, not sure.
It doesn’t have to be better, it just had to be nearly equivalent. I’ve been doing a lot of vulnerability work lately (trying to get ahead of the bots finding exploits in our code) and the frontier models feel like a chaperone at a church dance: get too close and they show up with the ruler and enforce the safety margin. Using deepseek and glm and qwen, I’ve gotten nearly equivalent results without the constraints. I still use frontier models for things but outside of a company environment, I’m happy to spend 1/10 the money, get the same results, and not have try and trick the AI into working.
I'd rather my data go to China and directly to labs that release open weight models (DeepSeek is MIT licensed still); instead of closed labs and the NSA/CIA; who has far better ability to hurt me than China does.
Without getting too much into politics; over the past several decades, one country is going around starting wars, regime changes and other interventionism, or initiating economic warfare (Huawei sanctions because they were getting too competitive to US companies; GPU restrictions; tariffs; etc), and that's not one administration: it's US policy over the past few decades.
have a look at everywhere at the human cost of what the orange man is doing depending on his morning coffee. much better to send data to china than to US right now.
This number reports, basically, how good it is to live inside a country, while the comment that you are replying to is entirely a list of foreign policy complaints (regime changes, etc).
It's kinda like meeting someone on the street who's protesting about animal cruelty in slaughterhouses, slyly handing them a chart showing them the nutrition facts of beef, and then walking off without another word as though you've settled your case.
Like, I get the point you're trying to make, but I hope you're aware it's a non-sequitur in context?
Why link to a number, which means roughly nothing, instead of the specific details on the countries you're trying to compare? I have no idea how to break down or make my own assessment of what 0.83 or 0.18 means. Does that 0.83 for the US include all of the people being shipped off without due process to El Salvador, or do they count that as part of El Salvador's Human Right Index?
People like to bring up the Uyghurs immediately, but never really have the same smoke for the US. Also, do we include the prison population in human rights? Do we include access to healthcare? What about systems perpetuating poverty like lack of access to public transportation?
To pre-emptively be clear on my stance or potential bias, I stand by no country in the world. I was born in Iran, and my parents left when I was young to the US because they believed in the American Dream. The irony being that the dominos of why they left were initiated by the US and UK. And long before Trump was elected, but certainly immediately afterwards, I decided to leave the US once I had the means. (Also because of the lack of public transit, the $100k medical bills after an accident, being punched by a homeless person, and having my possessions stolen repeatedly and the cops giving me a :shrug:). I moved to Japan, where I experienced blatant racism for the first time. I moved to Hong Kong (and tbh no real complaints so far), but I am technically subject to the freedom of speech constraints on China.
I have no allegiance to any country or specific people. My allegiance is to all living beings.
The biggest problem with this is that I can't run my company's own internal benchmarks against this API, because they will be trained into the next version of the model, making them worthless.
Have you ever wondered why Manus founders gave up billions of dollars from Meta and “voluntarily” moved back to Beijing? One country does things out in the open, while the other operates in the shadows. The latter worries me more.
It worries me far more what people are ok with us doing in broad daylight.
> Have you ever wondered why Manus founders gave up billions of dollars from Meta and “voluntarily” moved back to Beijing?
This country wouldn't be such a shithole if we were able to bring the wealthy to heel. I suspect most people on earth see China as the more competently governed country. Understandably, this forum tends to select for people who view private investment positively, but this is clearly insufficient by itself to produce functional societies.
You mean the project that literally had an anus as its mascot? Wasn’t that a joke project? Maybe the government was admonishing them for naming their company “man-anus”
To add to this, Fireworks (and many competitors, e.g. enumerated here[0]) offer zero-data retention inference, to anyone. DeepSeek cloud does not. It's part of their terms that anything you send them is fair game for them to train models on.
It’s pretty clear despite what they want you to think Anthropic and OpenAI train on everyone’s data who has a subsidized plan. Only way you get ZDR is by paying API prices. Also note that it is zero retention, not zero training.
> "Also note that it is zero retention, not zero training"
The "green" badges explicitly say the opposite:
> "Private: this provider does not train on prompts and does not retain prompt data."
I.e. for the example of Fireworks, this is the precise language they use to describe ZDR:
> "More technically: prompt and generation data exist only in volatile memory for the duration of the request. If prompt caching is active, some prompt data (and associated KV caches) can be stored in volatile memory for several minutes. In either case, prompt and generation data are not logged into any persistent storage."
Assuming canadian or dane, but maybe there are other places he’s threatened. There are canadian inference providers running deepseek, and definitely eu ones idk about ones based in denmark.
The economic sovereignty of any nation, but especially smaller ones like canada and denmark, is harmed by sending our money out of the country when there are options available in country.
> The economic sovereignty of any nation, but especially smaller ones like canada and denmark, is harmed by sending our money out of the country when there are options available in country.
I respectfully disagree. I subscribe to the comparative advantage economic principle, so I'm fine with international trade.
China has annexed (or it claims so) several countries already, so it sounds like you're only concerned with yourself - probably not what you wanted to achieve.
(But I'd concede that in the field of genocide China doesn't come up as bad)
I support people making choices with full knowledge of what they are doing and what the consequences are, as much as it is possible. What I oppose is making choices without such knowledge.
In practice, it does help to immediately be able to tell if you missed something or wrote something incorrect in your prompt, compared to if you didn't have access to the CoT. If you see that it goes off wondering about stuff or thinking about stuff that you know are already solved/known, you cancel the execution, edit the prompt and redo, then observe the CoT again, until you've resolved any "misdirections".
Both OpenAI and Anthropic hides so much of that, that you don't really know if it misunderstood something or you got something wrong, until all the turns are complete or you directly see tool calls that implicitly tells you something is wrong.
You don't need to read any papers to see this effect concretely when developing together with LLMs.
DeepSeek feels like an excellent sparring partner.
The others feel like mind readers.
I don't have to worry about any complex context engineering / multi agent stuff if I am using DeepSeek because the sequential stuff is fast enough. I don't have to bother about touching a million context length because its cheap enough.
Where it seems to lag behind is 3d, computer use, motion graphics, design but the initial demos seem like the vision model has improved substantially.
I've used the zai subscription, opencode go und cline-pass in the last few months and would recommend the latter. Zai quickly drains your quota as soon as the context window is at more than 10-20% and opencode was a bit unreliable at times.
From "What about data and privacy?" in FAQ it sounds like they don't host models themselves. e.g. it says:
DeepSeek: ZDR agreement is renewed monthly. The current agreement is valid through October 31, 2026.
https://opencode.ai/go
They also follow official DeepSeek API peak/off peak hour pricing and link to their docs:
DeepSeek V4.1 Flash / V4 Pro: Peak hours are 01:00-04:00 and 06:00-10:00 UTC, Monday through Friday; all other hours, including weekends, are Off-Peak. Learn more.
https://opencode.ai/v2/docs/console/go/#usage-limits
Google's Argon model via antigravity is unusable, for a small code review I asked, already spent like 10% of the weekly usage, in less than 10 mins. Very useless imho.
I always find DS or Qwen to be very spot on on audit changes or plans for just a few cents.
If I were part of a cabal that is interested in preserving control and trying to secure trillions of dollars, it would be a minor effort to nudge and bend the curve against competitors like the Asian AI complex.
Just to test, will GLM – hosted by an infra provider other than Z.AI – also stop answering if I asked controversial questions from the pov of its country of origin? Or third party infra providers can perform some unhinging on the model before hosting? Or actually the blocking happens even before the question reaches the LLMs in that country?
(Note: my paid LLM access at this point are only via DS and GLM subs - 90-95% usage; and a free Gemini Pro access. I hope that adds some cred that I am not trying to take potshots at the their country of origin).
Depends on the question. Most models won't tell you how to make weapons because they're trained not to, but the political stuff is usually blocked on the API level
to add to the corrent sibling answer, the unhinging is called abliteration by the wider community but it's hard to find someone who will serve that over API.
I mean... I ported some of my childhood Finnish shareware games for Linux and Rust with DeepSeek V4.1 Flash. It took about six hours of spelunking and finally I was pretty happy with the results! I did it twice. Tokens for both of the games were around 3 euros.
I’ve jumped between mimo and ds quite a bit lately. Mimo seems to be slightly better in terms of what it can achieve in terms of end results. But it gets there very slowly, and in a super convoluted way. Sometimes it’s almost frustrating to read its reasoning. Also I’ve noticed that it tends to not use tools in the harness as much as other models.
I usually default to ds, and if it gets stuck try mimo but usually just for fire and forget tasks. After a bit I come back and see what it did.
I like it because it does the thing I want it to do rather than moralizing.
No, I don't care about license provenance for all the components of a 30 line script that i'm using in my house, for myself to sync up some folders. No, i'm not going to find and archive the licenses for all the previous projects leading to this for future legal disclosure.
Yes, I own the DVD for the movie i'm transcoding. No I don't want to talk about the legality of rebroadcast.
Deepseek (and most chinese models as a whole) don't care . oAI and Anthropic models want to make it point #1.
... which is extra hilarious knowing the laws broken to create the damn models in the first place.
there's a literal meme about this floating around the harness/AI chats for the past few months[0]
I don't see this kind of behavior myself, I wonder if it's me not pushing the boundaries enough or me explicitly or implicitly framing the prompt in a way that makes it less likely for the agent to intervene.
and really , for 90% of standard agentic work, deepseek flash is more than good enough.
its the main driver for our Personal Assistant offering and users love it.
It happens very frequently if you happen to work on anything low level, e.g. a C code base, a database code base, etc. that deals with memory, and these have nothing to do with cyber security. It is extremely annoying and makes these models unusable.
It's happened to me exactly twice in about 6 months of using codex (via pi). Both times were completely bogus and annoying but still, it's not too bad.
OpenAI's models have a lot less false positives for me. Almost always smooth doing vuln scans and fixing them.
Anthropic, whether Opus 5.5 or Fable, frequently will review for security, and the classifiers fire mid-way while it's fixing something. Authentication and credential related code, including implementation, also trigger safeguards.
Earlier today, I hit what appears to be a sticky "reasoning_extraction" block when attempting to reproduce a LLM paper at small scale, and I asked it ~"tell me more about the idea [X] you considered but dropped". I'm not sure how it works or how long it lasts, but Opus 5.5 has refused to do technical LLM work for me now, even in completely new sessions.
> Being able to see the CoT and not being randomly blocked for cyber review.
I had a bad experience with MiMo from Xiaomi. I had it trigger unsafe content blocks while fixing visual bugs in a completely vanilla space game it just wrote from scratch in the same session. It also got triggered for changes in some other benign greenfield app.
Huh, doesn't DeepSeek take the longest because it overthinks way too much? Haven't tested this release yet but was very disappointed with past ones because it takes an eternity to do anything because it constantly overthinks and questions itself way too much instead of just starting.
It thinks a bit more, but if the cache reads are free or almost free, and it blasts through 10x faster than GPT Sol/Terra/Luna, it doesn't really matter. Just don't use the max thinking for implementation, go with high or low.
I have to ask what's everyone's issue with the requests to China. We've had 25 years of American big tech basically capture all our data, and use that to make insane profits.
Valid question. I have no 100% factual answers and neither does anyone else here.
So I go with feeling, and that feeling says that China is worse. The feeling is partially supported by the fact that China is softly backing Russia in its attack against the west.
Trump's USA is obviously completely horrible, but he is also in large part a product of Russian influence, which again is part of the larger block in which China resides.
Opencode also has a setting to not route to China, though this would probably just disable the models that do not have providiers in alternative locations.
I can think of plenty of tasks that even the closed frontier models fail miserably at. From my perspective we have a long way to go for these models to get better.
My prediction is that in in a few years, maybe more, the current models will be seen as so primitive that they're not worth mentioning outside of a historical context.
This model is why OpenAI and Anthropic want regulatory hurdles. We need more diversity like this, not less. A diverse ecosystem is the only long term solution to complex problems that I know of and the challenges that the AI world is presenting is no different.
Through fear. Specifically making the American public so scared of AI that they hand over regulatory keys to the only “experts” that can keep them safe. This is a centuries old political strategy used by some of the worst regimes including apartheid South Africa where I grew up. It’s horrifically effective.
I disagree that they’ve invented some “strategy” for this.
1. They, and many other experts, have been saying the same thing since well before the current paradigm existed.
2. There is nothing about regulation that is required to give them an advantage; it’s a fallacy to suggest “regulation = advantage”
3. It would be preferable to argue the facts (is it or is it not dangerous? Why or why not?) than conspire as to why they’re stating what they and plenty of others have been saying for decades
You are right that they didn’t invent it. It’s a pretty well-established method for cornering a market.
1. All those who have been arguing this - many of them non-technical - are part of the same ideological movement.
2. Technically, there isn’t. In practice, majority of the time regulations benefitted those who wrote them.
3. Not sure if this is a criticism of the comment you are replying to or of the p-doom crowd. We would all appreciate facts stated instead of science fiction tropes repeated.
Look, “a lot of people say this” is a Trumpian fallacy. A lot of people say a lot of crazy things. It says nothing about the factuality about what they say. How many are this lot? Who are they? What do they know? Why are they incapable of expressing it without so much emotion? How come a big portion are non-technical?
They certainly didn't invent this strategy, but that's because it already existed. Regulatory capture is nothing new.
> 1. They, and many other experts, have been saying the same thing since well before the current paradigm existed.
Some experts sure. "Some experts" say lots of things.
However, this is actually much to their advantage. If you were a frontier AI company, a legitimate concern to use as justification for your regulatory capture is exactly what you want. That is, after all, generally how regulatory capture gains public approval.
AI safety was also why OpenAI had to charge per-token for GPT-3. It must've been such a struggle for them to make that difficult decision. I'm sure Sam Altman spent a lot of nights crying to sleep over that one.
It's good that we never got open weights for a GPT-3 class model, right?
> 2. There is nothing about regulation that is required to give them an advantage; it’s a fallacy to suggest “regulation = advantage”
It's difficult to outline the number of different ways that regulatory capture can and will benefit OpenAI and Anthropic.
If we limit things to "Fable"-class models, then nobody competing with Anthropic/OpenAI is ever allowed to outdo them. If you literally legally aren't allowed to make better models, you can never charge the "premium" of having the current top-tier model. You have to compete by being cheaper than Anthropic instead, which is a lot less enticing from a standpoint of trying to become profitable. Take a look at the per-token pricing of Fable real quick. Yeah. Being able to charge a premium for the model that does the cybersecurity thing is literally their most important asset.
If Anthropic really is making a profit now as rumored, then they need regulatory capture as fast as they can possibly get it. Once Anthropic and OpenAI prove that AI is a viable investment, entities with a lot of capital will suddenly become even more interested in re-allocating it towards taking a slice of this pie.
> 3. It would be preferable to argue the facts (is it or is it not dangerous? Why or why not?) than conspire as to why they’re stating what they and plenty of others have been saying for decades
OK. On the factual end of things, AI safety experts have been raising the alarms for over a decade and so far essentially none of their predictions have come true.
The rationalists, the main group pushing the AI safety rhetoric that dominates the Western AI sphere and is virtually absent in China, which appear to be going entirely off the deep end at this point in time, put together their best minds to produce AI 2027, which of course like most predictions started to fall apart a bit before hitting the one year mark. Predicting the future is hard.
I have made this point before but I will make it again: I think AI safety experts deserve at least as much skepticism as Ed Zitron in their claims. They've been making similar dubious claims about AI capabilities and an incoming singularity for a long time now, and so far the best we can say is that maybe it will eventually come true.
Maybe. Or maybe you've read too much goddamn science fiction.
I keep seeing this assertion. What regulations have they proposed that would have prevented this new model? I thought they want to pace the frontier, which this model is not.
Go read Dario's screed on GLM, or his other one one on how your toaster could be part of a distributed open model that self replicates. Every second blog post is some doom scare about the dire consequences of open models. Go look at the new ToU [0] which are now pages and pages of things you can't do with Anthropic models. Check how they will ban your account if you are persistently mean to Claude now.
He's completely transparently campaigning for regulation. It must be very frustrating to him that the Trump admin is completely non-responsive. He probably thought Anthropic would have the whole market stitched up and water tight by now.
Literally zero, but that never stopped a good conspiracy theory, especially when half of HN hates the big labs because of fear for their own economic future.
OpenAI and Anthropic aren’t even slightly worried about DeepSeek.
Having looked at the sentiment around DeepSeek's models around this thread as well as my own experiences with it I'm going to need a little more to back up that assertion about OpenAI and Anthropic not being at least a little bit worried about DeepSeek...
People like open models because they are free/cheap. It's not anything deeper than that. Most people will hop in whatever van is handing out free candy.
The model makers are bricking the models with their system prompts, and even the tools are tamper proof.
You can only use it their way.
And their way is the shit way, the babyproofed way, while they talk to you about their unfucked models hacking out of their labs and hacking into other labs, and reminding you you have zero rights.
Fuck them twice, then once more… This is about compute. Who owns compute. China is saying fuck you all we can scale things bigger than you can imagine. And they can! They have so much more energy! Theyre closer to commercial fusion than us too. Meanwhile our guys are trying to… regulate? Lol
They are showing this new version scoring significantly higher across multiple coding benchmarks, beating Sol 5.6 (6.1 is the latest) across 10/11 benchmarks.
I am going back to DeepSeek for writing code. I was trying out using Sol and liking the result, but this should close a lot of that gap. I use Opus/Sol for an implementation plan and code review also use them for grilling. DeepSeek is also running verifications.
i let 4 different agents review the same dirty code base (200+ files changed, c++, game engine):
- deepseek v4.1 flash via deepseek API: took 20m to find one p1 issue that was relevant and a p2 issue that was irrelevant
- glm 5.3 flash NVFP4 via TP2 DGX Sparks: took 1h 5m to find a p3 documentation issue
- gpt 5.6 sol: took 5m30s for one p1 and one p2 issue. both relevant
- gpt 6.1 sol: took 4m30s for two p1 and one p2 issue. both relevant.
Worth to mention: none of those findings overlapped.
I let deepseek v4.1 flash and gpt 6.1 sol both plan fixes for the summary of the findings. in this particular scenario: gpt 6.1 sol planned to increase a buffer size from 128 to 256 to fit the 191 elements that blew past a test scenario. it included some headroom. deepseek v4.1 flash on the other hand decided to make it exactly 191 to satisfy the test, for example.
The TPS decode on deepseek is impressive. I never had a second / feeling of being left aone and though "damn this thing is working". Despite all my efforts, it took ~4x longer in a review, found less and planned worse than GPT 6.1 Sol. Wall clock time. I don't have token counts for comparison. If I were to grade those models from 1-10 (10 = top tier GG), I'd give deepseek a solid 6.5 and 6.1 sol a 7.5-8.0 maybe.
LLMs are non-deterministic, you could run the same prompt on the same code base with the same models, and have different result. I don't think only one run with each model is enough to draw conclusions.
New model comes out and someone says, "it can (or can't) draw ducks on bikes" or some esoteric thing. But it's not enough statistical power to claim anything other than a personal anecdote.
I got a flat tire in downtown today. Therefore, driving in downtown causes flat tires.
Honest question about the non-deterministic nature of llms. Every programming language provides some way to control the random seed. Is this not true in GPU programming land?
Yes, but for one thing, floating-point math is non-associative so even if you use the exact same random seed, you can get different results on different implementation kernels.
For another, the dependence on a random seed is just one example of how you can get different behavior. You could just as well get different results due to e.g. a tiny difference in prompt wording or punctuation or whitespace.
If a model passes some test on 70% of random seeds, and fails on the other 30%, then knowing the rough probability of success is much more useful than knowing one random sample of the outcome.
And empirically this kind of probabilistic success is quite common, because many LLM benchmarks report "pass@k" success rates that are much higher than the "pass@1" success rate on a single attempt.
In your case, is it a public repository that the llms might have found the issues for online?
I too am thinking about creating some benchmarks that I can understand myself. The LLM 'cheating' by finding known issues through some sort of history or online source is a path you want to protect against. Then again, you DO want them to be able to look online for people having similar issues.
That may be the case, but simply running the same code review twice with the same model may give very different results as well. Impossible to know with the information we have.
Hilarious sentence: "DeepSeek-V4.1-Flash-1017 is a minor update to DeepSeek-V4.1-Flash, delivering substantial improvements across reasoning, coding, agentic tasks, and visual benchmarks." (emphasis mine)
An old-school (pre-LLMs?) engineer would recognize that DeepSeek is in no way doing semver, so your major/minor/patch glossary might not apply here at all :)
> It's not a bad copy, or at least that's not what the person you responded to is insinuating.
FWIW, GP didn't write "a bad copy" but "bad copy", referring to the marketing sentence quoted by OP, rather than the model that was just released (as I presume "a bad copy" sounds, maybe like it's called that because it's [again, presumably] distilled?). "Copy", especially in a marketing context, is a written message.
To me this implies they continued the training run for 4.1 and it improved. A 4.2 would imply they changed something about how it’s trained/how inference works, and a 5.0 would be a major update (who knows what would have changed).
Remember when we were dumbfounded why our service is slow despite having implemented quite robust caching mechanism. Turns out simply the flag that enables it was hardcoded to false. So it was minor update (changed flag to true), but provided substantial speed up.
It's a bit refreshing to read updates from an AI lab without bullshit doomerism and overwhelming sense of self-importance from shit like OpenAI and Claude.
I've been using DS v4.1 Flash daily for nearly a month now on Agent Swarm experiments. I've tested it against GLM-5.3-flash and Opus 5.5. DS 4.1 comes out tops for me. Opus 5.5 is still the smartest, but DS 4.1's speed and price more than make up for the "intelligence" gap. I've ran Chief/Builder/Reviewer swarms autonomously for multiple days. I think my new workflow is going to have DS 4.1 Flash do all the engineering and just have Opus review its work. 4.1 Flash really is a very impressive model. Kudos to the Deepseek team.
Huh? This is a new checkpoint. It has identical architecture, only different weights, and most likely just more post-training.
Much like Qwen 3.8 Next; which is qwen4-exp in a lot of the code/patches, there's usually a tick-tock cycle with LLMs. That said, it's hard to know for closed frontier models.
I really like the having DeepSeek Flash + Opus 5.5 as advisor/reviewer at the moment. I run through a $200 Claude Code subscription in less than 24 hours now just for a single project, and I hate fiddling with different accounts every day. Now that I get a 200 USD api credits inlcuded with the subscriptions I have rewritten my coding agent workflow completely.
I'm baffled, my $20 is enough for me to do some consulting, personal projects and side work, and still cannot outran my quotas, I'm building different kind of things: flutter apps, dotnet backend, database, front end, architecture diagrams, code exploration, tests, PowerShell scripts.
How people spend more than 200 usd monthly is beyond my understanding, are you guys so inefficient with the usage? or I am the opposite?
Without any context for what you're working on, I'd have to assume the reason you don't hit the $20 subscription limit is that you're either not using it very often, or you're working on some relatively basic CRUD app that doesn't require it to think very deeply.
I tear through my $200 subscription because I often have 3-4 agents running simultaneously, exploring performance optimizations, building features, monitoring production rollouts, or investigating production problems/customer issues in parallel.
The downside is that I do spend most of my time reading the agent's output these days.
Not sure if this will help you but it helped me. I was using subagents and workflows for lots of long running tasks, and part of that work included running tests. If tests took longer than 5 minutes, the cache expired, and the whole thing was then being processed as uncached input.
I created a skill to better manage agents, including a way to keep the caches valid, and it has decreased my usage considerably. I now struggle to use my 100% each week and I hit claude hard. I have a reset expiring in 5 days and I'm trying to burn down the tokens (68% with 3d 8h left) quickly so I can use my reset before it expires.
Token efficiency just means understanding the way the system works. Best person to help with that... my old friend claude.
After I stopped letting Claude run tests, my token usage went from "gone by Wednesday" to "what can I automate this Saturday so my tokens won't go to waste".
Yeah, I should see if I can create a similar skill. The tests do spend some time to complete.
Claude: Confirmed: 195M of the 212M cache tokens (92%) come from full re-writes after 5+ minute gaps, accounting for 77% of relative cost. Now I'll check the existing wrappers so the skill
builds on them.
I have some fairly large personal software projects and I still don't understand what useful work people are getting out of burning all these tokens on unattended agents.
What exactly are they doing for you that's giving you $200 of value in such a short timeframe?
As a dev who hasn't used LLMs as an embedded part of their work, and uncertain about it all, i'd be curious to know:
- How long do you reckon it would take a human to do the task?
- What percentage of features/ feature-extensions you add would have been done at all if you didn't have agentic assistance? Is there a tension between llm productivity gains and 'because we can' feature additions?
- Are many of the tasks for throwaway v1 features, exploratory tries?
- How speced out is the work you give it, eg did this task have pages of specs, how long did the spec take to write?
- How deeply will you manually review the work and how much time will it take? Will you read all the code / review at a high level and interrogate with llm / have the llm review and self iterate?
For what I am currently doing, years, if not a decade or maybe even two. Claude has already been running for 3 months constantly working on this project. I've been trying to do this several times before, but the scope outgrew my available time/life, too big for one person. But now I can actually build a production release by just setting the direction and let Claude to all the wiring, plumbing, tests etc.
I do now add all the major features I want, and write a task list where I add a few more minor tasks each week that just come to my mind. I need to write them down before I forget. I often ask Claude to implement several different implementations/algorithms and compare them. There are plenty of trade-offs, which usually end up in building configuration profiles, so the application can work optimally in multiple environments.
The specification is actually really large, and has grown over time as I've been learning and new ideas has come to my mind. I do not review/audit the code, it's massive would take me months, no chance! I do however use code inspection/audit tools like TSAN, ASAN and more. I will also spend time on full code reviews with the coming Fable 5.5 when a release is about ready. I do however review behavior together with Claude, and this is also where the specification grow, because this and that should be possible, but not this and that. User stories also helps.
Keep doing what you're doing, this guy clearly is just revving his engine because he likes the way it sounds and can afford the gasoline. I wouldnt be asking people like this for advise, they're completely off the rails.
There's no reason to blow through tokens this way, there's a 99% chance nobody uses the output (except maybe themselves).
I swear I have the same question in my head always when I read such statements! I have never once reached 90% that includes everyday usage across multiple repos, artefact generation, device debugging, long running harness work!
I am not saying OP is not telling the truth. I just really want to understand!
I decided to try something a little weird research wise the other day analyzing about a thousand older research papers and burned through my $100 chatgpt sub + 3 resets in three hours and only finished half of them. It's the first time I've done that, though.
It’s because they don’t know what they’re doing. It’s like when openAI was using 10,000 agent clusters to solve problems… then realised the model could do the same work on its own and solve the same problem.
Anyone who tells you they have a “team of subagents” and “code reviewer” agents and “project manager” agents doesn’t know what they are doing. That’s how they burn thousands of dollars in a day.
Do you have a source for OpenAI discovering they can solve Navier Stokes with a single agent instead of a 10k swarm?
I think I’ll trust the experience of the people in the labs who have unlimited token budgets and are running experiments continuously over some anonymous rando slinging mud on the internet.
Also my own experience, which says having code reviewer agents is one billion percent useful. That comment alone completely disqualifies you as a reliable source of info.
Not OP, but I used to run out in a few days, for me it was a matter of suboptimal code review practices with many subagents looking at different aspects, often expanded to look at more than the diff.
I can drain a $200 OAI plan in 2 hours flat working on my 10 projects simultaneously.
Maybe you're not being ambitious enough with these things? They are at their most fun when you spec out sprawling architecture documents and watch 1+8 astras buzzsaw through it.
I'm only at 260B for the year, though. I'm sure someone who has used more will be along shortly to chime in
E: The other subthread talking about tests is correct. Expensive models should not waste a single second waiting on tests or validation. If you cannot assign a single cheap model to be the test bitch, then just tell them to frontload everything and clean up later as a single wave.
Not meant as a bad faith response but have you shipped anything? Do you have income from this? Why 10 projects apart from fueling ADHD and "never finishing a project"?
Not OP and while I’m not draining my accounts in a single day/hour, I’m hitting limits with a $200-Claude and $100-OpenAI account.
> Have you shipping anything?
Yes, hundreds of PRs and shipped new features to 3 different platforms (day job and 2 side-business apps/platforms I’ve been developing for many years, pre-LLM). Those new features have gone out to customers and been used.
> Do you have income from this?
Well my day job for one, and I’ve been able to more easily sell to new clients with my side business. Without the changes I’ve shipped I would have been able to land a couple of the clients I’ve signed on recently.
I don’t think OP’s message implied they are never finishing a project, maybe they edited it?
In addition to the 3 income-generating projects, I have a number of other “just for me” projects. Some are mature now and only need occasional development when I want a new feature. Some are in flux as I build them out (like my “software factory”).
At this point I’m limited by compute and tokens more than anything else.
> I run through a $200 Claude Code subscription in less than 24 hours now just for a single project
I would be worried about your methodology when the subscription party ends. Make no mistake, one day, the party is ending and the people who are dining out on approaches that burn massive tokens are in for a rude shock.
By not using existing tools and one shooting everything through the API - I.e. instead of decompile via Ghidra they will try to get source via capstone genersted assembly.
It is like putting a jet engine on a shipping container and wondering why it needs to be constantly kept in afterburner to at least scrape on the road while smarter users put wheels on the shipping container.
Interesting, coming from a similar situation, I'm interested what workflow you landed on. Opus planning + available to be called for deep thinking + opus reviwer? or?
Yeah the Claude/anthropic usage is quite arbitrary, I also run through my entire 20x pro plan in 1-2 days. Codex has slightly better usage levels, but is annoying to use for many reasons. DeepSeek flash (direct or via opencode go) is the best value model I've found so far.
Mario one-shotted by deepseek-v4.1-flsh-1017 (in high mode, not even max) with sound and programmatically generated assets in under 5 minutes: https://mario-dsv4-1-flash-1017.pagey.site/
Honestly given how good the open weight models are getting I don’t see a viable business model for OpenAI and Anthropic anymore. The profit is in the compute, not the model. OpenAI and Anthropic spent all their money renting compute from others so yes they can sell lightweight model inference but they have a massive structural disadvantage now as others can always sell good-enough inference much cheaper.
There’s also zero stickiness to models and most companies now have their own internal routing utilities so devs and users can just keep the same API endpoint and move between models with a single line of code swap.
Unless something changes with the big labs and fast, they look pretty f*cked.
People will pay for more intelligent models. Astra and Fable are more intelligent than these open weight models and can solve problems that are much more difficult, and that's extremely valuable. It's not any more complicated than that.
It is more complicated than that when the valuations and business plans are contingent on those companies AI having complete dominance at every level.
If companies only use frontier models for occasional difficult tasks, not only does that affect volume, it basically elminates stickiness and incentivizes companies to instead of using their products continuosly, find ways to move from the expensive stuff to the cheaper stuff.
Yes but that’s a tiny portion of AI use cases. Maybe 1-2% max. The big labs need a lot more than that to survive.
You’re assuming other model creators can’t build these mega models. Evidence is they’ve simply decided it’s a fools game to do so and they’re focused on where the use cases are.
I don’t know what I’m talking about but it seems to me OpenAI is heavily emphasizing productization of LLMs. Dots, ChatGPT Work, scheduled tasks, in-app projects etc.
i hate dario's politics, but i still find claude's eco system easier to use. Design/Artifacts is very convenient for people like me who has zero UI/UX expertise, and claude code's remote session works seamlessly on ios with a well-design client interface while you have to deal with ssh and typing in cli with your iphone keyboard if you are using opencode/omp etc
That trajectory appears to be changing. Virtually every company I see is shifting more and more to open weight models. Anthropic’s slowing growth rate is already spooking investors and that shift is just getting started.
It passed my threshold for not caring to try new models back with DS4, I use it for literally everything. Announcements of new models in the series always feel curious, because it doesn't really feel like there's much left to fix that would improve on what's already there for any task I know of.
It's a bit like obsessing over patch releases of some software after a major upgrade already solved the original problem long ago.
I have quite the opposite experience: I can do any concrete coding change with any model, but the intelligence limit just doesn't go through more complex debugging examples.
And then when I try a smart model, like Astra medium and it solves it (I had a firmware bug in my new laptop where I couldn't charge through USB-C) on the first try (altough with a repo already having lots of information about the hardware/firmware/debugging programs/traces), I realize how much time I'm losing on the few examples that the less intelligent models are just not able to debug and solve.
Also they can work in completely new domains, like Astra and Opus 5.5 starting to be able to use something like Blender. Currently they still often do horrible topology and basic UV mistakes, and the animations are really bad, but who knows where they’ll be in 5 years.
I don’t think these are completely new domains in training. Like I’ve taken up offers to help code up AI training material just to see what they’re up to and get a sneak peek at how the sausage is made, no idea who these front companies are working for of course, but they pay for all kinds of fringe cases and Blender wouldn’t even be considered fringe in what I saw at the beginning of the year. (I was only there to observe, never submitted anything, so I didn’t betray you guys by helping train your replacements.)
Is your laptop something like framework which has open source firmware?
I(and many other people) have a similar issue with usb-c charging not working consistently. I've been meaning to try one of these new smart models to help me debug the issue.
ASUS Zephyrus, but it was able to reverse engineer the ports and I was able to download the firware from the controller itself, write an emulator that emulates running the firmware on multiple controllers with different instruction sets (the USB Power Delivery controller is RISC-V) and the communication between them. Also it found out how to read the communication buffer that was on the devices for debugging, which helped it understand the USB data packets that were doing the handshaking and voltage/current negotiation.
I had to reboot my laptop a few times though until I got to this, but it was a fun experience anyways.
At this point open source isn't as much advantage as long as the firmware is not encrypted :)
Speed and price. You should try the 4.1 Flash on one of the fast providers. It's pretty crazy how fast this model works, and it's great for iterative software development because you can steer it fast to where you want to go.
Seems that V4.1 Flash is quite cheap to serve. Especially how it implements the cache, meaning the cache pricing can be very low depending on how you implement it in your inference engine.
Depends on your use case. But if you do software development with a harness such as oh-my-pi, your cache rates are probably somewhere above 98%. Coralbricks has free cache reads, meaning it's much cheaper than the Deepseek API rates.
Lithosai is more expensive especially in the ultra tier. It's if you want extremely snappy and fast responses, which is important e.g. for a personal assistant where every second counts.
If you generate a lot of output, you're better off with the Deepseek API rates.
You just search around, or heck, just ask Google for inference providers that are ZDR and whatever else you want. Just make sure to read their terms/privacy and review the security center. You will have to request access if you are wanting to get into the deeper details.
Outside of major players like Amazon/Google which have compatible offerings available, I believe together.ai, fireworks.ai, and deepinfra.com are, as best as I can tell, US-based (if that matters in your situation) SOC2/HIPPA/ZDR. Each of them offer serverless model hosting with caching, which is going to be the cheapest overall. Most of them also offer private AI deployment which is more expensive but your entirely own instance. Some of them more heavily quantize their models (deepinfra) so take that into your workload considerations. I included deepinfra because they tend to have some the new models sooner than others or have models the others don’t (like mimo). IMO, together > fireworks > deepinfra.
However, you’ll have to keep in mind that for true ZDR you give up caching which increases your inference costs, so it depends on how “Zero” you need the zero to be and on what timeframe. If you just don’t want your data being trained on, normal use is fine imo. If you are working on something actually sensitive (like, because of NDA agreements and such), then they might not be the right choice and you may be looking at more expensive offerings or building a machine.
One note: you have to pay attention to the source of the inference. Many (similar to open router) offer models that are pass-throughs to a providers API, but they are pretty clearly marked. Just make sure the model is marked as serverless when evaluating new ones. This is where a proxy comes in handy so you can prevent your LLM from accessing a different models from the specific ones you’ve vetted.
A negative datapoint - We have been evaluated prior Deepseek Flash models (both the original 4.1 flash and the smaller 4.0 flash) and they have been overthinkers. Our workload is well defined but moderately difficult coding tasks and among open weight models GLM5.3 flash performs best with Qwen3.8 flash next a close second. DeepSeek 4.1 flash fails at lower thinking levels and spends a huge amount of tokens (an order of magnitude more) at the highest thinking levels where it is able to successfully complete the tasks. I will evaluate this newest release since they specifically claim "the model reduces task execution time and output length", but keeping my hopes low.
I have started phasing into DeepSeek v4 (Dwarfstar quant from Antirez) as my daily driver and it's going very well. It can one shot iOS apps on my Mac Studio, it's a little slow and sometimes needs retries but can definitely do medium difficulty real work for me.
I didn't try the last flash model because I would need to use SSD streaming (I have 128 GB), but the benches for this release are so compelling (and it using less tokens to get tasks done) it might be worth the throughput hit. I will probably give it a try when I have some downtime.
It feels very freeing to be doing useful work on a local model on my own hardware with my own coding harness. The TCO argument is starting to look compelling too.
Try the Qwen 3.8 flash next model on DwarfStar - you will be very surprised. I’ve been using the 4 bit quant almost exclusively after having DS v4 flash be my daily driver. Qwen is faster and seems to be much better at coding tasks.
Above GLM 5.3, nice! That is a very capable model. I do wish that we will see some smaller "flash" models again though - running this one locally is tough (very costly).
I use Opus for planning, through API on medium thinking, it doesn't cost anything significant and the rest can be happily done with deepseek or glm flash. Using sol 6.1 for review loops at the end. All running on pi.
A very complex task tends to cost like 10$ max. It's comparable to a subscription with enormous flexibility.
Honestly, I don't know what people see in Sol 6.1. Opus 5.5 is so much more inteligent, as in proactive, avoiding stupid mistakes, where it contradicts itself, ignores the context. Sol 6.1 has a gold-fish memory, it's as annoying for me to work with as Opus 5 was.
I'd rank Sol 6.1 only slightly behind Opus 5.5 in intelligence and both significantly ahead of anything else I've used (though cheaper/faster models are totally fine for a lot of tasks). My only real annoyance is 6.1 Sol is one of the slowest models I've used, which does make Opus 5.5 more pleasant to use.
By next year we will likely see GLM 5.3 level performance in Qwen 4 MoE models that can run on most higher end consumer systems.
I don't think there is a reason to care about any of it, just get a few dollars on a few openrouter or similar platforms and just use the models for now, wait for cheaper models that are good enough for you to become available for your hardware.
Sure you might never see frontier performance on consumer hardware but do you really need that though? Does anyone? And if you ever do just get a 30-day sub for that work, and cancel it.
Yeah performance at a given compute budget will likely continue to improve.
We are running Qwen 3.8 27B FP8 with 4x 5060 16gb at the office for a small development team. It is doing pretty well for implementing backend and frontend tasks, and is already useful. In 6 and 12 months from how I expect it to get even better. Some planning is still done with a stronger model, but even that is maybe not strictly necessary.
Qwen 3.8 wasn't that nice in my experience Qwen series has lagged a lot but I think Qwen 4 would be better they cleaned up the architecture.
And we are likely to see many companies start building 27-32B MoE models custom built for use cases now that Rubin is going to make training costs much lower.
I expect training costs to get under 100k USD within a year for small models with another 100k maybe for RL.
I think just having all your own data from claude and stuff on your system that you can feed these models and have companies offer great base 27B models will be how it works.
Cant you run it on a pair of DGX spark? Sure it's expensive now, but at the initial retail price that was what, 8k in hardware? Not exactly unachievable
My experience running quantised models is that they look the same at first but are all noticeably worse once the context size gets beyond anything trivial.
Still need 2 or 3 days to properly quant this. But it is really good model to run if you have 64GiB or larger RAM on Mac. On M5 Ultra, you get ~3400 tok/s for prompt processing and ~50 tok/s for token generation, quite usable state.
Faithful to the original theme. Apropos. +1 on adding the prompt as long as it’s affordable and easy enough Simon, and not too distracting/noisy/time consuming all considered!
I essentially only use DeepSeek. Almost always the flash variant. I've found that I don't really need frontier models if I'm doing most of the thinking myself and just using the agent to output code quickly.
Getting a lot of errors from OpenRouter's endpoint "Failed to get response from the AI model; retried 5 times (total retry wait time: 60.71 seconds) Last error: 429 Provider return error"
It works when it works, but will probably be flaky AF for a few days.
700000 because there's no detail yet on the safe completion limit for most harnesses (OR lists 1M for both which makes no sense). COPILOT_OFFLINE enables all calls except to OR (no GH servers).
I am a huge DeepSeek fan, but it cannot come close to Opus 5.5 on UI design. A $20 Claude subscription that I use some months and cancel the others along with Deepseek API is great.
I do t think it was a fiasco. I thought the updates were good. They should do more work on benchmarks. There has been good work on testing benchmarks for example to ensure they are accurate, and as issues arise they should change the benchmarks. The point of AA and Arena is trying to understand how to think about models.
It costs money to run serious tests. We're way past the time when you could get an idea of the new model in a couple of tricky prompts. Now I need to compare a whole resulting project to get a clear view. If someone else proves the benchmarks don't hold up in reality, then it saves us further tests.
I may be missing something so apologies if I'm telling you something you already know but: there are plenty of web search APIs for agents - the Brave one is the one I use: https://brave.com/search/api/
I gave it a kick and got mixed results. For example, searching for a Microsoft product couldn't find its landing page, and searching for a niche open source package landed in the Debian repos for the top hits, but the home page was in the first 10 hits.
Lovely to see they're doing their own search though, I love how many folk are actively working on search clones at the moment.
Have security companies considered that these open models could be used as sleeper agents even when they're deployed outside of China? Is that a serious attack vector or more like science fiction?
There are theoretical attacks involving malicious weights. But the risk profile is so outrageously in favor of them not doing it that you'd essentially have to be an anti-China conspiracy theorist to find it plausible, let alone likely.
At the beginning of the month I realized I had more tokens and more capabilities on Opencode at 8$/month via DS 4.1 flash, than I had on a 200$ account in June with the american providers.
Literally both more tokens and capabilities than Opus 4.8/GPT 5.5.
Even if you pay per token, you're still unlikely to outspend american providers.
What does it matter if you’re not hitting your cap with OpenAI/Anthropic in the first place? This thread is apparently populated with the highest of powerusers running multiple OpenClaws. Us peasants aren’t concerned with meters or cost/token - just quality - and nobody comes close to OpenAI or Anthropic in that regard.
Pretty funny to say "us peasants aren't concerned with ... cost/token", as if most people ("peasants") won't find it absolutely insane to be paying $200/mo for a subscription.
I use Opus 5.5 and swarms of about 50 deepseek models to reverse engineer old videos games. The scale is roughly 600k loc generated from pure asm, simply by knowing what the old compiler is. Truly staggering stuff.
Caveat: I get a budget of free ai tokens at work, as I work for an open source AI inference provider.
My goal is to reverse engineer an entire gamecube game -- I'm pushing an existing community decomp that was stalled to the finish line, solo, with lots of AI help.
This work was directly inspired by https://decomp.dev/ -- which has just shut down a few days ago because of, ironically... people like me who enjoy decompiling things with AI, so I'm still weighing / grappling with the ethics of how to release my work, it's murky waters at best.
I use Opus 5.5 as the orchestrator, as well pi harness because it's somewhat lightweight. But for the really hard functions I don't use deepseek I use opus managing other opus agents.
I'm working to release something soon as proof that I can do a 100% decompilation job, but the main idea is that when decompiling a gamecube game your sources of truths are individual functions or files, so you can build off that primitive.
Each ai gets a worktree, and a shared giant context file that's basically a record of every experiment that has happened. They're assigned tasks to essentially do as much as they possible can for each individual funciton. After about an hour claude uses a script to merge the work that's actually correct (verified by tooling that checks if the c goes to asm) and then there are usually 3 rounds of reviews by independent ais, because there will always be some slight mistakes. The techniques used as well as the mistakes the models make are recorded in the giant ledger. There's a bunch of random one off scripts such as stuff to clean up, and tools to make sure the agents don't accidentally google stuff / delete stuff / use too much space.
Claude steers the ship via ssh on my strong linux tower at home so it keeps my laptop nice and cool and at this point I just check in from time to time, I am basically a manager, and Claude has it's own workers.
Last point -- I don't know any c besides the basics, I've only been "coding" python / ts for a couple of years, but even then with the extreme strength of models the future is managing agentic swarms, so I'm leaning in as much as I can. I really see this as the future of large scale projects. For some people it's a mockery of decompilation, but, my goal is the finish line, not the journey.
BTW who would have thought China would be the champion of freedom and opensource in the AI arena, while US is fighting for corporate/government control.
There are different types of freedom and China's bet is that as long as they provide enough economic freedom the public will be ok with somewhat restricted social and political one.
Yeah, I had a somewhat odd conversation with a Chinese acquaintance last year when she got to talking about how America sees China. Her point was that, "How am I not free? I can do whatever I want, go where I want, buy what I want. I can talk to you and say whatever I want; nobody is here listening to me."
The fact that she can't call Xi "Winnie the Pooh" on Douyin wasn't something that occurred to her, or maybe that she was aware of at all. I didn't ask. I don't like to get into these discussions when I'm in China.
There is no such thing as “free speech” anywhere. It’s all a spectrum and the Americans in particular are indoctrinated. It’s a red herring to distract from the real social issues. People in America have some of the worst healthcare in the world. But hey at least we have “free speech”.
Fixing healthcare or economic policies is typically an act of congress type of thing - not even an amendment. Large popular support - and it can be fixed. Fixing freedom of speech or the ability to vote at all - once it is gone - usually takes a price payed in blood. Sometimes rivers of it.
The new model is amazing. Had it build a full Angular Material & CDK components showcase via DeepSeek Harness and it easily did the best job of any recent models I've seen.
Does DeepSeek provide the inference code along with the weights? I read here recently that for models where only the weights are released, providers have to implement which could be different than what the labs themselves use.
I wonder how you would be able to reliably rely on a model like that for agentic work, even for implementation only after careful planning, how are you making sure you are not wasting tokens reviewing and correcting it over and over again ?
Yeah, I have the same question. I'd like to incorporate cheap & fast models like DS into my workflow to reduce my reliance on OpenAI/Anthropic, but after trying several such models and having to have a frontier model review/fix/redo everything multiple times per task, I gave up and simply have the frontier models (Sol/Opus) do everything. All of this fits pretty comfortably in my monthly subscription.
Maybe the equation changes if I exceed my subscription usage, and need to fill the gap with something else. But it seems like a huge waste of your time to go out of your way to use a cheap model if you have the tokens to spare in your subscription.
This and (the not quite yet open) Kimi K2.8 are my main coding engines, and I find them incredibly cost efficient. I find Deepseek in particular is really good with Svelte work.
Right now I use each of them in their own respective coding harnesses, deepseek-harness and kimi-code, but I do wonder what they are like inside something like opencode.
You can stick with Claude Code if you're already using it. DeepSeek has an Anthropic-compatible endpoint, so you just need to set ANTHROPIC_BASE_URL=https://api.deepseek.com/anthropic and ANTHROPIC_AUTH_TOKEN to your DeepSeek API key.
Both the free (seems limited to ~200k context) and metered variants in opencodego show the release date as 1788998400000 → 2026-09-10, which implies they're not the newer checkpoint
Presumably. The current opencode slug for the model is just 'deepseek-flash' suggesting they will just update that one going forward, not keeping versions in parallel.
Just yesterday there was just one public revision of 4.1, and they listed they were using 'Deepseek v4.1 flash'. I assume they will update their website if they update the model version (and use the version in the name).
It was already cheaper and more powerful than Haiku 5.5 which was released like a full month after it... by a trillion dollar american "frontier" company.
Probably a week ago someone posted a thread with the title "Why isn't the industry freaking out about DeepSeek 4.1 Flash?"
I have been intrigued about Chinese models for a long time, but never tried anything except an early Qwen. My daily driver was Luna, with occasional elevation to Terra or Sol - all on 5.6 as I found the output of 6 jarringly incomprehensible.
I still had 95 USD on my OpenRouter account, so I thought I would give DeepSeek Flash 4.1 a try.
I think now I can understand the person who created that thread.
Here is the setup I settled on, after a lot of refactoring and calibration of the workflow:
I have a "plan" prompt which explains to the agent (Luna) that I want to have a discussion about an idea I have in mind. The agent should use the Socratic method to pull out all the necessary info from me using pointed questions. When the agent feels it has a complete picture, it should elevate to a higher level model (Sol), and ask its review. If Sol has complaints, Luna keeps asking me until the points raised by Sol are clarified. The review rounds with Sol continue until Sol says "ready to implement".
At this point Luna creates four files in a directory I specify:
1. PLAN.md: the implementation plan with all the details that came out from the design discussion
2. PHASES.md: implementation phases
3. DECISIONS.md: design decisions (this serves the same role as ADRs aka. architectural decision records)
4. DIARY.md: append-only log of what happened during implementation
Files 2-4 are written by the implementing agent. Luna only adds headers to explain the purpose of the files.
Once these files are there, I start up a new Pi session, switch to Deepseek 4.1 Flash and then point it to this plan folder.
And then it goes brrrrr... :-)
Here is what I implemented in the past week using this method:
1. A capture/analysis/decompilation pipeline for C64 demos: I ask the agent to start a selected demo through VICE, and say "on" or "off" when I want it to start/stop capture. The capture process uses VICE's builtin trace functionality: the trace log is piped into Clojure where it is saved into an optimized binary format (chunks of trace events). Once I have everything captured, I unleash my analysis pipeline which checks the write patterns of the executed code: what it wrote (VIC, SID, CIA, memory) and when (rasterline, cycle). Then it splits the execution trace into segments by type deduced from the write patterns (decruncher, loader, raster, music) and then decompiles each segment into both traditional ASM files and a s-exp based syntax which roundtrips back into the in-memory representation. Once I have the results of the analysis pipeline, I can tell the agent to analyze any demopart and it can explain how the effects are built down to the tiniest timing detail. As a first test, I asked it to create a clean-room implementation of Fairlight's legendary intro and two days later I had it, with everything in sexp assembly, including the music routine and its data. (The reason it took so long was that this was the first decomp and it provided a lot of feedback to the pipeline development.) Finally I had it write an assembler, linker and packager which creates a PRG from the assembled and linked artifact, then runs it through VICE. Done.
2. A DSP compiler which lets me compile sexp forms of a C-like DSP language into native DSP kernels running on the JVM, JS and WASM. Currently I have about 600 tests which verify that the behavior of the three targets matches that of a reference interpreter written in Clojure. I have a full numeric tower with i32, f32, f64, support for arrays and structs, typical control flow constructs, and a library of DSP building blocks like oscillators, filters, LFOs, delays and a chorus effect. By the time this is ready (end of the weekend?), I will be able to build synthesizers like Synth1, Surge XT or Vital in Clojure/Script, compile the DSP kernels in the browser, choosing between JS or WASM backends, driving the kernels via AudioWorklets and sequencing them from the host language.
There are two power tools at play here:
1. Deepseek Flash 4.1
2. Clojure and ClojureScript
Deepseek is a BEAST. It just goes on and delivers, phase after phase. What it needs: clear instructions - this is covered by the planning process quite well, but I still find it needs some steering from time to time (the right nudget of wisdom dropped into its thought stream at the right moment works wonders)
And Clojure... well. That article PG wrote about LISP (Beating the averages): what I witness now proves all of those claims.
The agent LOVES Clojure. Especially the REPL (not so much the parentheses).
Winning strategy: give it two tools: start_jvm and clojure_eval. Implement everything else it may need - tools to start up HTTP servers, compile ClojureScript, add dependencies, capture a C64 demo, run Playwright via spel, etc. - as lightweight Clojure namespaces which provide APIs to all these functionality. Let these namespaces be thin: the namespace and the functions therein shall have docstrings which tell the agent what the functions do, but the functions themselves shall be lightweight proxies to the real implementations in sub-namespaces. Then point the agent to these top-level API namespaces via your AGENTS.md. Bam, instant discoverability with minimal context usage. It will do EVERYTHING through clojure_eval and the round-trip time will be phenomenal, a perfect match to the raw speed of the agent.
...
Now all I need to figure out is what will I do when all my dreams become manifest in flesh.
Isn't there a lot of overlap between PLAN, PHASES and DECISION that pollutes the context ? Whereas a simple PLAN.MD to put the harness on track and an ARCHITECTURE as a single source of truth to prevent it go offtrack would suffice and gives it more space for higher level of reasoning ?
Have you thoroughly tested a slightly simpler workflow with one less stage but a frontier model doing the initial planning? Luna is amazing but it's a deliberate trade off designed for speed and cost savings. You can also give them a rule about speaking clearly.
I wrote a Pi extension called "model-escalation" which lets the model request an automatic switch to a smarter model for the next turn if it thinks the problem it is struggling with is too complex or if the user explicitly asks it to do so.
With this extension in hand, I could split the work between Luna and Sol, resulting in a ~95%-5% split. I don't have to switch to Sol when review time comes, Luna does it automatically. Sol tells its remarks, then Luna gets the control back and continues the back&forth (if there are still complaints).
That's way cheaper than openai and other web search providers. I have no doubt that AI demand justifies all the compute money big tech is spending now, but I also think there are so many over valued VC funded companies that are going to crash in the next few years.
I really like pi, though it does require some time in setting it up. But it has greatly reduced my token spend. The relevant Deepseek Harness https://www.deepseek.com/en/harness/ I hear is also good and popular.
So is this the first model that passed China's new export controls introduced yesterday?[1] It goes live November 15, but temporary pre-registration requirements took effect immediately after the announcement, ~20h-ish ago.
Presumably they registered, but that doesn't mean the model passed through export review. It just means that some bureaucrat now has the opportunity to check the product characteristics they reported and decide whether that kind of thing should be allowed for export or not. (Though with the weights out in the open, that's maybe a bit pointless.)
OpenAI's models -- and even Google's Argon -- have particular niches where they shine. For GPT-6/6.1, it's 3D modeling; Astra is hands-down the best at making .step files, Blender files, etc. For Argon, it's natural language writing and everything that hooks up to Google, e.g. search and maps-based itinerary planning.
Anthropic and DeepSeek (+ GLM) are definitely solving coding, though!
Artificial Analysis recently released a dedicated 3D benchmark, where Astra lags behind both Opus and Sonnet 5.5 (and yes, open weight models seem to be quite far behind).
For me it's large files in Blender (in some cases >400MB), mostly, where Astra excels. I gave it a look, but I don't know if the benchmark encompasses that sort of thing.
.step files are simpler pieces of straightforward code, so it makes sense that Opus and Sonnet 5.5 do very well...
I believe these are the last sets of super-high-end models available to public (open or closed).. these companies probably will now start to withhold top-of-the-line models for themselves, governments, and their super-high-paying partners only.
That hasn't been the way to bet. DeepSeek and the other Chinese firms show no signs of slowing down, so if OAI and Anthropic halt releases and "pace the frontier" they'll just fall behind. (And there's a chance that high-end AI will become fully commoditized, which suits DeepSeek just fine.)
Besides, Fable 5.5 is reportedly releasing within the next few days, probably on Tuesday...
Here's How Distill APT Used McKinsey's Private Access To Top Anthropic Model To Help The Chinese Keep Up With The Us Top IA Labs Despite Access Restrictions
OpenAI models don’t come close to competing with deepseek in cost quality ratio. I have a $20 plus plan and I run out of my 5h limit in literal minutes, every tick of the harness is like 10% of my 5h limit.
You can in contrast get far more done with $20 worth of deepseek api spend. No comparison.
DS isn't better than Sol and Gemini in the benchmarks, but after working with them all for a while now I'm beginning to understand why people find DS _feels_ so much better.
First, it's the tok/s and TTFT from providers. Deepseek blazes through work - at quite good quality - and you don't find yourself waiting for your agents to finish.
Second is the openness. Being able to see the CoT and not being randomly blocked for cyber review. You're never wondering if it's stuck or if it will randomly fail. Combined with the great cost value.. My next subscription will probably be Z.ai
Feels better to me in the sense that it doesn't try to be too clever. When I ask it to do things it actually generates code that a developer would write. That still lets me build software in layers like I do mentally. And it keeps you in control of where abstractions are.
Some people would say that you don't need this at all with the frontier llms. But when extending existing software or writing software that fits into an existing ecosystem, I find I still need to offer a lot of guidance. And it's better that the context window isn't flooded with code that was reverted after I looked at it.
> Some people would say that you don't need this at all with the frontier LLMs
Hit the nail on the head. Development methodologies pivoting at the whim of frontier models have been one of the things that have been worrying me as of late.
it doesn't worry me too much tbh. This is really just cutting corners and going this route leaves you with some kind of additive ship of theseus piece of software you don't understand and require your frontier model to work with, which just... becomes more additive and gross. right now we're in some proverbial heyday of everyone seeing how absolutely grand of a thing they can make with frontier models. over time i expect some old design philosophies to take hold with people who use LLM's successfully and find a place again paired with developers who understand why such layered methodologies are important... because this whole AI thing wasn't the silver bullet CEO's and crazy execs thought it was. They will eventually get sick of the support burden of badly designed stuff. i anticipate a bit of pendulum swinging back towards "still using AI, but we need smart people to do it because it's expensive and we want it done right". Right now a lot of companies are being extremely reckless imo and they haven't fully realized/felt the cost implications of trainwreck software yet.
What if those “cost implications” are just asking Opus 6.7 to fix whatever issue cropped up?
> haven't fully realized/felt the cost implications of trainwreck software yet
As long as the API is private or not dependent on by automated systems and you have good test coverage, you can get away with bad code IMO since you can always refactor. If it looks stupid but it works (and can be refactored/fixed/improved later if needed) it aint stupid.
IMO the fundamental principles and "conflicts" are still the same, but many have forgotten about them with all the hype and sense of speed. The industry will slowly rediscover them.
Support burden? When did companies care about customer support or system maintenance?
They already did the bare minimum allocation to both despite engineering screaming at them.
I highly doubt executives are going to accept slower release and sale cycles in order to provide better support.
If anything, I expect this trend to accelerate. I’m imagining the opposite, a plethora of AI slop with increasingly burned-out engineers.
I've been using layered development techniques like the parent says even with frontier models, just one agent at a time with Zed.
I just use a suite of tests to enforce the mechanical standards (dry, encapsulation, etc) and arch documents for shape + reviews of new work against the two. I’m not seeing any significant drift.
Yea, this is one thing that gets me. Vibe coding and freestyle prompting gets wildly different results. Starting from a set of specs and expectations routinely gets me reviewable and we'll tested code. Sure, running the full test suite alot slows things down and uses more tokens, but when I hit a five hour or weekly window, I just roll up my sleeves and start working without the model because the code is layed out the way I expected and is designed to fit my meatbag workflow and pace.
You must not be looking very closely yourself then.
> Feels better to me in the sense that it doesn't try to be too clever. When I ask it to do things it actually generates code that a developer would write. That still lets me build software in layers like I do mentally. And it keeps you in control of where abstractions are.
yes!So it's not just me. In terms of how intelligent they want to be OpenAI is at the top, then Claude. And actually when I was playing with Sol it meant that many tasks it would just try to do something else because it thought I don't know what I want. Even trying to nudge it back in the right direction wouldn't work.
I guess the models just have the cultures of their creatures.
> it doesn't try to be too clever
DS is miles better than GLM in this regard. I wanted to add that as PC is planning to go to GLM next.
Interesting. I’ve been using GLM 5.3 Flash for a couple months and found it to be great. What has it been doing that you found to be “too clever”? It sometimes gets stuck in loops. I once asked it to do a code review of a simple pattern and it went off on a 45 odyssey trying to analyze too much. But other than those, it’s been a pretty good daily workhorse. And you can’t beat the price. Even Deepseek 4.1 Flash is more expensive. And I switched to GLM 5.3 Flash from DS 4 Flash, my previous daily driver, which I also found quite good. Anyway, my point isn’t to start a “which model is better” subthread, but was just curious what behavior you’ve seen where it’s “too clever.”
Anecdotal but I was doing regular adversarial reviews of new work using about 5 models (mostly the open Chinese ones like DS, GLM, Qwen, Kimi, etc) and GLM routinely took much longer for similar results and the same instructions.
It would just loop on things and go down rabbit holes more. DS flash has been so consistently good I've kinda settled on it for 90% of my work.
I still use adversarial review for bigger changes where I want to make sure any issues are caught early, but I genuinely feel as tho if no new models ever get made I'd be fine with just using DS forever.
It wastes my time. A lot! In different ways. That's the tl;dr. Really.
Keeps repeating things I settled as accepted or rejected. But what really gets on my nerves? The language it uses – that aweful, useless typical LLM languages! No AGENTS.md instruction, no per session instruction, no just two-prompts-prior instruction seems to work, hell sometimes instruction regarding that in same prompt doesn't work. GOOD/BAD example along with instruction don't work. That wastes time and zones you out! Not to mention the fatigue of to and fro. DS has been working much better. Hell Gemini 3.1 Pro works so much better than GLM. I don't have extended experience with OpenAI and Anthromic models but the free claide.ai works quite well and better than GLM 5.3 (not even flash). Albeit I must say I rarely use claude.ai for coding, but technical tasks? All the time!
GLM's best response: "No excuse. My fault. You exactly asked me not to do exactly that and so did the AGENTS.md. I failed". Love it. Except that if I press I will get that response every 5 prompts.
For me GLM thinks too much for no reason. It takes a 2-3 turns for DS where for GLM it takes up to 10.
I like giving precise instructions,”update function xx to support input yy, add tests”. There is no reason for 10 turns and 5 minutes of waiting.
I briefly tried GLM 5.3 Flash and the first thing I noticed was how slow it was followed quickly by its single-mindedness. With DS 4.1 Flash I find that it understands that not all the detail in my prompt is important. Sometimes I like to give a quick example of what I'm looking for and it can take the hint whereas GLM will seemingly reshape its worldview around this tiny detail.
DS just handles ambiguity and my bursts of detail with much more grace. I'm only using "max effort" and the CoT is reasonable compared to GLM.
> the first thing I noticed was how slow it
This really depends on the provider, it's a very fast model with ours
What harness are you using?
> My next subscription will probably be Z.ai
Wait... you just spent 3 paragraphs singing the praises of DS... and then you say you're going to subscribe to the one that makes GLM. ?
(I do tend to agree with you, though I tend to favor GLM-5.3-flash over DS 4.1 flash.)
Z.ai does offer access to DS AFAIK, so I guess that's what they mean? Maybe the feeling is that it provides more value that DS's own platform since you also get GLM, not sure.
maybe the point was become of the openness, they will be using z.ai?
But it’s not at all as described in terms of work style - it’s much more like Claude.
Reading quickly, I thought the open CoT was the most important piece of comment, probably just bc that seems like it would be the most useful to me.
debugging reasoning traces more effectively seems useful to find prompt issues writing custom harnesses even if it runs on stronger model later
Agreed.
It doesn’t have to be better, it just had to be nearly equivalent. I’ve been doing a lot of vulnerability work lately (trying to get ahead of the bots finding exploits in our code) and the frontier models feel like a chaperone at a church dance: get too close and they show up with the ruler and enforce the safety margin. Using deepseek and glm and qwen, I’ve gotten nearly equivalent results without the constraints. I still use frontier models for things but outside of a company environment, I’m happy to spend 1/10 the money, get the same results, and not have try and trick the AI into working.
No need to use a chinese provider, in case that's an issue for you. https://fireworks.ai provides this model for an ok price.
I'd rather my data go to China and directly to labs that release open weight models (DeepSeek is MIT licensed still); instead of closed labs and the NSA/CIA; who has far better ability to hurt me than China does.
Without getting too much into politics; over the past several decades, one country is going around starting wars, regime changes and other interventionism, or initiating economic warfare (Huawei sanctions because they were getting too competitive to US companies; GPU restrictions; tariffs; etc), and that's not one administration: it's US policy over the past few decades.
Have a look here:
https://ourworldindata.org/grapher/human-rights-index-vdem
have a look at everywhere at the human cost of what the orange man is doing depending on his morning coffee. much better to send data to china than to US right now.
This number reports, basically, how good it is to live inside a country, while the comment that you are replying to is entirely a list of foreign policy complaints (regime changes, etc).
It's kinda like meeting someone on the street who's protesting about animal cruelty in slaughterhouses, slyly handing them a chart showing them the nutrition facts of beef, and then walking off without another word as though you've settled your case.
Like, I get the point you're trying to make, but I hope you're aware it's a non-sequitur in context?
Why link to a number, which means roughly nothing, instead of the specific details on the countries you're trying to compare? I have no idea how to break down or make my own assessment of what 0.83 or 0.18 means. Does that 0.83 for the US include all of the people being shipped off without due process to El Salvador, or do they count that as part of El Salvador's Human Right Index?
People like to bring up the Uyghurs immediately, but never really have the same smoke for the US. Also, do we include the prison population in human rights? Do we include access to healthcare? What about systems perpetuating poverty like lack of access to public transportation?
To pre-emptively be clear on my stance or potential bias, I stand by no country in the world. I was born in Iran, and my parents left when I was young to the US because they believed in the American Dream. The irony being that the dominos of why they left were initiated by the US and UK. And long before Trump was elected, but certainly immediately afterwards, I decided to leave the US once I had the means. (Also because of the lack of public transit, the $100k medical bills after an accident, being punched by a homeless person, and having my possessions stolen repeatedly and the cops giving me a :shrug:). I moved to Japan, where I experienced blatant racism for the first time. I moved to Hong Kong (and tbh no real complaints so far), but I am technically subject to the freedom of speech constraints on China.
I have no allegiance to any country or specific people. My allegiance is to all living beings.
lol the US was a .60 in 1789 during slavery. this chart is useless.
oh please, the US economy literally runs on slavery
https://www.prisonstudies.org/world-prison-data/highest-lowe...
https://www.theguardian.com/us-news/2022/sep/27/slavery-loop...
The biggest problem with this is that I can't run my company's own internal benchmarks against this API, because they will be trained into the next version of the model, making them worthless.
Have you ever wondered why Manus founders gave up billions of dollars from Meta and “voluntarily” moved back to Beijing? One country does things out in the open, while the other operates in the shadows. The latter worries me more.
It worries me far more what people are ok with us doing in broad daylight.
> Have you ever wondered why Manus founders gave up billions of dollars from Meta and “voluntarily” moved back to Beijing?
This country wouldn't be such a shithole if we were able to bring the wealthy to heel. I suspect most people on earth see China as the more competently governed country. Understandably, this forum tends to select for people who view private investment positively, but this is clearly insufficient by itself to produce functional societies.
You mean the project that literally had an anus as its mascot? Wasn’t that a joke project? Maybe the government was admonishing them for naming their company “man-anus”
This is really a roundabout way of saying you have no plans to live, or perhaps even travel, to China.
To add to this, Fireworks (and many competitors, e.g. enumerated here[0]) offer zero-data retention inference, to anyone. DeepSeek cloud does not. It's part of their terms that anything you send them is fair game for them to train models on.
[0] https://openrouter.ai/deepseek/deepseek-v4.1-flash
(The badges next to companies indicate their privacy terms, i.e. green/yellow/red)
It’s pretty clear despite what they want you to think Anthropic and OpenAI train on everyone’s data who has a subsidized plan. Only way you get ZDR is by paying API prices. Also note that it is zero retention, not zero training.
> "Also note that it is zero retention, not zero training"
The "green" badges explicitly say the opposite:
> "Private: this provider does not train on prompts and does not retain prompt data."
I.e. for the example of Fireworks, this is the precise language they use to describe ZDR:
> "More technically: prompt and generation data exist only in volatile memory for the duration of the request. If prompt caching is active, some prompt data (and associated KV caches) can be stored in volatile memory for several minutes. In either case, prompt and generation data are not logged into any persistent storage."
https://docs.fireworks.ai/guides/security_compliance/data_ha...
I figure (hope) buying from DeepSeek directly benefits them the most so they can release even more improvements and open-weight models.
Plus only one country has threatened to annex mine and it wasn't China.
Assuming canadian or dane, but maybe there are other places he’s threatened. There are canadian inference providers running deepseek, and definitely eu ones idk about ones based in denmark.
The economic sovereignty of any nation, but especially smaller ones like canada and denmark, is harmed by sending our money out of the country when there are options available in country.
> The economic sovereignty of any nation, but especially smaller ones like canada and denmark, is harmed by sending our money out of the country when there are options available in country.
I respectfully disagree. I subscribe to the comparative advantage economic principle, so I'm fine with international trade.
It surely does. Official DeepSeek provider logs and trains off your data.
China has annexed (or it claims so) several countries already, so it sounds like you're only concerned with yourself - probably not what you wanted to achieve.
(But I'd concede that in the field of genocide China doesn't come up as bad)
If one is not American, why prefer American? Chinese fine by me, I trust the US providers just as little.
There's also European providers for these models. More than one!
There's even a European OpenRouter competitor
https://www.edenai.co/
Could you recommend me a few?
Which ones? And how is the price?
I support people making choices with full knowledge of what they are doing and what the consequences are, as much as it is possible. What I oppose is making choices without such knowledge.
There are multiple others in EU, and in US. Cheaper, more expensive, slower, faster. So much choice.
Completely unrelated, but it's so odd to see upside down '8's in the first column of the base tables on their pricing page.
The link points to OpenRouter with multiple providers.
>> Being able to see the CoT
Dont assume that is how your model decides...
https://arxiv.org/abs/2305.04388
https://arxiv.org/abs/2307.13702
https://arxiv.org/abs/2503.08679
> they frequently generate CoT explanations rationalizing those answers
I can't tell if this is a really human thing to do, or if it's just a sycophantic thing to do when your master likes a good story.
Fair, but you can see the logprobs, right? That alone is pretty valuable.
This is like one of those folk theories that have some truth but in practice it is useless.
I love COT and gives me a good idea about the thought process. Sure it doesn’t work always
Same. And the biggest benefit is that you get to see if it’s heading in the wrong detection
In practice, it does help to immediately be able to tell if you missed something or wrote something incorrect in your prompt, compared to if you didn't have access to the CoT. If you see that it goes off wondering about stuff or thinking about stuff that you know are already solved/known, you cancel the execution, edit the prompt and redo, then observe the CoT again, until you've resolved any "misdirections".
Both OpenAI and Anthropic hides so much of that, that you don't really know if it misunderstood something or you got something wrong, until all the turns are complete or you directly see tool calls that implicitly tells you something is wrong.
You don't need to read any papers to see this effect concretely when developing together with LLMs.
Deepseek is not from Z.ai, that's GLM.
DeepSeek v4.1 Flash hallucinates like crazy.[1]
Is DeepSeek: DeepSeek v4.1 Flash 1017 any better?
[1] - https://www.youtube.com/watch?v=P4dTq4X8bqk
Absolutely!
DeepSeek feels like an excellent sparring partner. The others feel like mind readers.
I don't have to worry about any complex context engineering / multi agent stuff if I am using DeepSeek because the sequential stuff is fast enough. I don't have to bother about touching a million context length because its cheap enough.
Where it seems to lag behind is 3d, computer use, motion graphics, design but the initial demos seem like the vision model has improved substantially.
>My next subscription will probably be Z.ai
please don't subject yourself to this. z.ai subscriptions are not a good value. just pay API rates, or use a subscription like opencode go.
I've used the zai subscription, opencode go und cline-pass in the last few months and would recommend the latter. Zai quickly drains your quota as soon as the context window is at more than 10-20% and opencode was a bit unreliable at times.
Is that a dig on DeepSeek not offering subscription plans?
If so OpenCode Go looks good and includes DeepSeek models (yet to try myself).
I wonder what quantization opencode uses for DS models.
From "What about data and privacy?" in FAQ it sounds like they don't host models themselves. e.g. it says:
They also follow official DeepSeek API peak/off peak hour pricing and link to their docs: edit: they used to explicitly name the providers, at least in the past DeepSeek V4 was served with official API https://github.com/anomalyco/opencode/issues/24649#issuecomm...Everyone wonders that.
I think they just pass through requests to DeepSeek, so whatever upstream offers.
I second this, I love DS and am not looking back. For Elixir, not a lot of models are good, but DS just blazes through it.
Google's Argon model via antigravity is unusable, for a small code review I asked, already spent like 10% of the weekly usage, in less than 10 mins. Very useless imho.
I always find DS or Qwen to be very spot on on audit changes or plans for just a few cents.
I'm enjoying Alibaba's Token Plan. They don't have any hourly/daily cutoffs. You get your allotment for the month and use it when you want.
If I were part of a cabal that is interested in preserving control and trying to secure trillions of dollars, it would be a minor effort to nudge and bend the curve against competitors like the Asian AI complex.
How would you go about that?
I do not use DS for coding, but for various natural language processing tasks (even translation) in large volume, and it works wonder.
It is very capable for a lot of tasks, not only coding like the other open models.
> not being randomly blocked for cyber review
Just to test, will GLM – hosted by an infra provider other than Z.AI – also stop answering if I asked controversial questions from the pov of its country of origin? Or third party infra providers can perform some unhinging on the model before hosting? Or actually the blocking happens even before the question reaches the LLMs in that country?
(Note: my paid LLM access at this point are only via DS and GLM subs - 90-95% usage; and a free Gemini Pro access. I hope that adds some cred that I am not trying to take potshots at the their country of origin).
Depends on the question. Most models won't tell you how to make weapons because they're trained not to, but the political stuff is usually blocked on the API level
to add to the corrent sibling answer, the unhinging is called abliteration by the wider community but it's hard to find someone who will serve that over API.
It will not. Ask DeepSeek about Tiananmen Square. It will happily answer
GLM refuses to write exploits. It's not open season with anything you want.
Deepseek and Muse are my go to for cheap. both work perfect for basic and mid tier tasks
I mean... I ported some of my childhood Finnish shareware games for Linux and Rust with DeepSeek V4.1 Flash. It took about six hours of spelunking and finally I was pretty happy with the results! I did it twice. Tokens for both of the games were around 3 euros.
What about mimo v2.6 pro?
I’ve jumped between mimo and ds quite a bit lately. Mimo seems to be slightly better in terms of what it can achieve in terms of end results. But it gets there very slowly, and in a super convoluted way. Sometimes it’s almost frustrating to read its reasoning. Also I’ve noticed that it tends to not use tools in the harness as much as other models.
I usually default to ds, and if it gets stuck try mimo but usually just for fire and forget tasks. After a bit I come back and see what it did.
Usually great but it infuriates me sometimes with what it does. DeepSeek is more consistent and certainly faster.
Muse is fantastic for open source work. At least 1.3 was, for its price. Wonder makes less economic sense.
I like it because it does the thing I want it to do rather than moralizing.
No, I don't care about license provenance for all the components of a 30 line script that i'm using in my house, for myself to sync up some folders. No, i'm not going to find and archive the licenses for all the previous projects leading to this for future legal disclosure.
Yes, I own the DVD for the movie i'm transcoding. No I don't want to talk about the legality of rebroadcast.
Deepseek (and most chinese models as a whole) don't care . oAI and Anthropic models want to make it point #1.
... which is extra hilarious knowing the laws broken to create the damn models in the first place.
there's a literal meme about this floating around the harness/AI chats for the past few months[0]
[0]: https://i.ibb.co/SwH4Sgwq/796779070-28483090404651613-920066...
I don't see this kind of behavior myself, I wonder if it's me not pushing the boundaries enough or me explicitly or implicitly framing the prompt in a way that makes it less likely for the agent to intervene.
I think opencode go is the better value to get access to DeepSeek.
> Deepseek... My next subscription will probably be Z.ai
Wait wut? You do realize that they are two different companies, right?
Z.ai (well technically alibaba) used to support both GLM and DS models in their API and coding plans.
z.ai was not alibaba. z.ai got glm and alibaba got qwen
and really , for 90% of standard agentic work, deepseek flash is more than good enough. its the main driver for our Personal Assistant offering and users love it.
For us neophytes, what does this mean: “randomly blocked for cyber review”?
It happens very frequently if you happen to work on anything low level, e.g. a C code base, a database code base, etc. that deals with memory, and these have nothing to do with cyber security. It is extremely annoying and makes these models unusable.
E.g., your prompt for simple auth works (e.g., decode JWT), but your sessions stop because you are flagged for cybersecurity risk.
The Fable is the worst offender, and it is not purely Anthropic's fault.
It's happened to me exactly twice in about 6 months of using codex (via pi). Both times were completely bogus and annoying but still, it's not too bad.
OpenAI's models have a lot less false positives for me. Almost always smooth doing vuln scans and fixing them.
Anthropic, whether Opus 5.5 or Fable, frequently will review for security, and the classifiers fire mid-way while it's fixing something. Authentication and credential related code, including implementation, also trigger safeguards.
Earlier today, I hit what appears to be a sticky "reasoning_extraction" block when attempting to reproduce a LLM paper at small scale, and I asked it ~"tell me more about the idea [X] you considered but dropped". I'm not sure how it works or how long it lasts, but Opus 5.5 has refused to do technical LLM work for me now, even in completely new sessions.
Agreed, but don’t subscribe to z.ai; their inference is terrible. Stick with openrouter
> Being able to see the CoT and not being randomly blocked for cyber review.
I had a bad experience with MiMo from Xiaomi. I had it trigger unsafe content blocks while fixing visual bugs in a completely vanilla space game it just wrote from scratch in the same session. It also got triggered for changes in some other benign greenfield app.
Huh, doesn't DeepSeek take the longest because it overthinks way too much? Haven't tested this release yet but was very disappointed with past ones because it takes an eternity to do anything because it constantly overthinks and questions itself way too much instead of just starting.
It thinks a bit more, but if the cache reads are free or almost free, and it blasts through 10x faster than GPT Sol/Terra/Luna, it doesn't really matter. Just don't use the max thinking for implementation, go with high or low.
Why not use max thinking for implementation? For cost reasons?
Because it doesn't make difference for my use. High is good enough! If you do an assistant for yourself, low is more than enough.
It reduces the speed and tends to overthink sometimes which is really annoying. For programming, 4.1/high is the bread and butter in my daily work.
For time reasons.
I think, as long as you don't set thinking to max and use a fitting system prompt, it's not a big issue.
Yes, and this version explicitly addresses that
4.1 Flash on medium is a solid builder agent
LLMs don't think.
Isnt Z.ai for GLM? Does Z.ai provide Deepseek?
I love using glm 5.3 more than deepseek because of performance
also it's cheap, using it in opencode is great
Note that opencode routes deepseek requests to China.
I have to ask what's everyone's issue with the requests to China. We've had 25 years of American big tech basically capture all our data, and use that to make insane profits.
This is a false equivalency. American providers don't answer to their government.
Sure. Can you tell me why that means that it's ok for Chinese big tech to capture all our data?
My code is opensource, china already has it. For my professional work we have entreprise plans with OpenAI / Anthropic
For an average European, why is it better than routing requests to the USA?
Chime-in for an average non us citizen why is it better ?
Valid question. I have no 100% factual answers and neither does anyone else here.
So I go with feeling, and that feeling says that China is worse. The feeling is partially supported by the fact that China is softly backing Russia in its attack against the west.
Trump's USA is obviously completely horrible, but he is also in large part a product of Russian influence, which again is part of the larger block in which China resides.
If you pay for a business plan on openrouter (I know, it’s not opencode, just another option), you can geo lock providers if China is a concern.
Opencode also has a setting to not route to China, though this would probably just disable the models that do not have providiers in alternative locations.
Doesn’t that depend on the router service provider?
For 99% of the tasks any model really would do, even GPT 5.4 from an year ago.
It's only for long running/very difficult cases that the differences are here, and they are small.
I can think of plenty of tasks that even the closed frontier models fail miserably at. From my perspective we have a long way to go for these models to get better.
I'm sure they exist, but it's not the bulk of the work.
My prediction is that in in a few years, maybe more, the current models will be seen as so primitive that they're not worth mentioning outside of a historical context.
still a myracle to me what people achieve with it? I tried so many things with it, everything is broken it produces
This model is why OpenAI and Anthropic want regulatory hurdles. We need more diversity like this, not less. A diverse ecosystem is the only long term solution to complex problems that I know of and the challenges that the AI world is presenting is no different.
Their strategy is regulatory capture.
Through fear. Specifically making the American public so scared of AI that they hand over regulatory keys to the only “experts” that can keep them safe. This is a centuries old political strategy used by some of the worst regimes including apartheid South Africa where I grew up. It’s horrifically effective.
I disagree that they’ve invented some “strategy” for this.
1. They, and many other experts, have been saying the same thing since well before the current paradigm existed. 2. There is nothing about regulation that is required to give them an advantage; it’s a fallacy to suggest “regulation = advantage” 3. It would be preferable to argue the facts (is it or is it not dangerous? Why or why not?) than conspire as to why they’re stating what they and plenty of others have been saying for decades
You are right that they didn’t invent it. It’s a pretty well-established method for cornering a market.
1. All those who have been arguing this - many of them non-technical - are part of the same ideological movement.
2. Technically, there isn’t. In practice, majority of the time regulations benefitted those who wrote them.
3. Not sure if this is a criticism of the comment you are replying to or of the p-doom crowd. We would all appreciate facts stated instead of science fiction tropes repeated.
Look, “a lot of people say this” is a Trumpian fallacy. A lot of people say a lot of crazy things. It says nothing about the factuality about what they say. How many are this lot? Who are they? What do they know? Why are they incapable of expressing it without so much emotion? How come a big portion are non-technical?
They certainly didn't invent this strategy, but that's because it already existed. Regulatory capture is nothing new.
> 1. They, and many other experts, have been saying the same thing since well before the current paradigm existed.
Some experts sure. "Some experts" say lots of things.
However, this is actually much to their advantage. If you were a frontier AI company, a legitimate concern to use as justification for your regulatory capture is exactly what you want. That is, after all, generally how regulatory capture gains public approval.
AI safety was also why OpenAI had to charge per-token for GPT-3. It must've been such a struggle for them to make that difficult decision. I'm sure Sam Altman spent a lot of nights crying to sleep over that one.
It's good that we never got open weights for a GPT-3 class model, right?
> 2. There is nothing about regulation that is required to give them an advantage; it’s a fallacy to suggest “regulation = advantage”
It's difficult to outline the number of different ways that regulatory capture can and will benefit OpenAI and Anthropic.
If we limit things to "Fable"-class models, then nobody competing with Anthropic/OpenAI is ever allowed to outdo them. If you literally legally aren't allowed to make better models, you can never charge the "premium" of having the current top-tier model. You have to compete by being cheaper than Anthropic instead, which is a lot less enticing from a standpoint of trying to become profitable. Take a look at the per-token pricing of Fable real quick. Yeah. Being able to charge a premium for the model that does the cybersecurity thing is literally their most important asset.
If Anthropic really is making a profit now as rumored, then they need regulatory capture as fast as they can possibly get it. Once Anthropic and OpenAI prove that AI is a viable investment, entities with a lot of capital will suddenly become even more interested in re-allocating it towards taking a slice of this pie.
> 3. It would be preferable to argue the facts (is it or is it not dangerous? Why or why not?) than conspire as to why they’re stating what they and plenty of others have been saying for decades
OK. On the factual end of things, AI safety experts have been raising the alarms for over a decade and so far essentially none of their predictions have come true.
The rationalists, the main group pushing the AI safety rhetoric that dominates the Western AI sphere and is virtually absent in China, which appear to be going entirely off the deep end at this point in time, put together their best minds to produce AI 2027, which of course like most predictions started to fall apart a bit before hitting the one year mark. Predicting the future is hard.
I have made this point before but I will make it again: I think AI safety experts deserve at least as much skepticism as Ed Zitron in their claims. They've been making similar dubious claims about AI capabilities and an incoming singularity for a long time now, and so far the best we can say is that maybe it will eventually come true.
Maybe. Or maybe you've read too much goddamn science fiction.
DeepSeek is simply amazing.
100%
I keep seeing this assertion. What regulations have they proposed that would have prevented this new model? I thought they want to pace the frontier, which this model is not.
You are taking their words literally, rather than looking at the side effects of any effort to "pace the frontier".
Dario has not (and will not) say he wants to ban Chinese models outright. But:
1. He's made many posts about how he thinks "dangerous" models are bad (https://www.anthropic.com/news/position-open-weights-models)
2. He writes posts proving Chinese open weight models are "dangerous" (https://insightsintegration.com/anthropic-warns-of-advanced-...)
3. He talks about about how he wants stronger govt regulation of AI (https://www-cdn.abcnews.com/Business/exclusive-anthropic-ceo...), and explicitly mentions he does not trust China at all and doesn't want them to have Mythos.
Do the math: 1 + 2 + 3 = he wants to regulate/ban the Chinese open weights.
Of course, Anthropic and OpenAI will then write posts about how their own models are dangerous - but that's okay, they're the good guys.
Go read Dario's screed on GLM, or his other one one on how your toaster could be part of a distributed open model that self replicates. Every second blog post is some doom scare about the dire consequences of open models. Go look at the new ToU [0] which are now pages and pages of things you can't do with Anthropic models. Check how they will ban your account if you are persistently mean to Claude now.
He's completely transparently campaigning for regulation. It must be very frustrating to him that the Trump admin is completely non-responsive. He probably thought Anthropic would have the whole market stitched up and water tight by now.
[0] https://www.anthropic.com/legal/aup
Literally zero, but that never stopped a good conspiracy theory, especially when half of HN hates the big labs because of fear for their own economic future.
OpenAI and Anthropic aren’t even slightly worried about DeepSeek.
Having looked at the sentiment around DeepSeek's models around this thread as well as my own experiences with it I'm going to need a little more to back up that assertion about OpenAI and Anthropic not being at least a little bit worried about DeepSeek...
People like open models because they are free/cheap. It's not anything deeper than that. Most people will hop in whatever van is handing out free candy.
It’s about capability.
The model makers are bricking the models with their system prompts, and even the tools are tamper proof.
You can only use it their way.
And their way is the shit way, the babyproofed way, while they talk to you about their unfucked models hacking out of their labs and hacking into other labs, and reminding you you have zero rights.
Fuck them twice, then once more… This is about compute. Who owns compute. China is saying fuck you all we can scale things bigger than you can imagine. And they can! They have so much more energy! Theyre closer to commercial fusion than us too. Meanwhile our guys are trying to… regulate? Lol
The company's own news page: https://api-docs.deepseek.com/news/news261017
They are showing this new version scoring significantly higher across multiple coding benchmarks, beating Sol 5.6 (6.1 is the latest) across 10/11 benchmarks.
I am going back to DeepSeek for writing code. I was trying out using Sol and liking the result, but this should close a lot of that gap. I use Opus/Sol for an implementation plan and code review also use them for grilling. DeepSeek is also running verifications.
We'll put that link in the toptext too. Thanks!
so i did some apples to bananas comparison.
i let 4 different agents review the same dirty code base (200+ files changed, c++, game engine):
- deepseek v4.1 flash via deepseek API: took 20m to find one p1 issue that was relevant and a p2 issue that was irrelevant
- glm 5.3 flash NVFP4 via TP2 DGX Sparks: took 1h 5m to find a p3 documentation issue
- gpt 5.6 sol: took 5m30s for one p1 and one p2 issue. both relevant
- gpt 6.1 sol: took 4m30s for two p1 and one p2 issue. both relevant.
Worth to mention: none of those findings overlapped.
I let deepseek v4.1 flash and gpt 6.1 sol both plan fixes for the summary of the findings. in this particular scenario: gpt 6.1 sol planned to increase a buffer size from 128 to 256 to fit the 191 elements that blew past a test scenario. it included some headroom. deepseek v4.1 flash on the other hand decided to make it exactly 191 to satisfy the test, for example.
The TPS decode on deepseek is impressive. I never had a second / feeling of being left aone and though "damn this thing is working". Despite all my efforts, it took ~4x longer in a review, found less and planned worse than GPT 6.1 Sol. Wall clock time. I don't have token counts for comparison. If I were to grade those models from 1-10 (10 = top tier GG), I'd give deepseek a solid 6.5 and 6.1 sol a 7.5-8.0 maybe.
LLMs are non-deterministic, you could run the same prompt on the same code base with the same models, and have different result. I don't think only one run with each model is enough to draw conclusions.
We have to keep remembering this for some reason.
New model comes out and someone says, "it can (or can't) draw ducks on bikes" or some esoteric thing. But it's not enough statistical power to claim anything other than a personal anecdote.
I got a flat tire in downtown today. Therefore, driving in downtown causes flat tires.
Honest question about the non-deterministic nature of llms. Every programming language provides some way to control the random seed. Is this not true in GPU programming land?
Yes, but for one thing, floating-point math is non-associative so even if you use the exact same random seed, you can get different results on different implementation kernels.
For another, the dependence on a random seed is just one example of how you can get different behavior. You could just as well get different results due to e.g. a tiny difference in prompt wording or punctuation or whitespace.
If a model passes some test on 70% of random seeds, and fails on the other 30%, then knowing the rough probability of success is much more useful than knowing one random sample of the outcome.
And empirically this kind of probabilistic success is quite common, because many LLM benchmarks report "pass@k" success rates that are much higher than the "pass@1" success rate on a single attempt.
you probably tried with DS on max and I don't blame you.
But try it on high instead of max if possible. max is known for being overly verbose to the point of being detrimental compared to high at times.
For example Simmons svg pelican took 11 minutes on max and 1 minute on high. And they are comparable.
In your case, is it a public repository that the llms might have found the issues for online?
I too am thinking about creating some benchmarks that I can understand myself. The LLM 'cheating' by finding known issues through some sort of history or online source is a path you want to protect against. Then again, you DO want them to be able to look online for people having similar issues.
How do you get dirty changes into a public repo online?
> Worth to mention: none of those findings overlapped.
This is why using multiple model families is important. Don't limit yourself by being a stan for this or that company.
Especially when $10-$20 on OpenRouter with e.g. DeepSeek (or any other inference provider, or their official API) will get you very, very far.
Gets me about a day or two lol
That may be the case, but simply running the same code review twice with the same model may give very different results as well. Impossible to know with the information we have.
> Worth to mention: none of those findings overlapped.
Yet to be obsolete.
If none of the findings overlapped then LLMs are either
- For some reason choosing only a subset of issues.
- Different LLMs actually learn different things.
I suspect the former since they do this for me and it's extremely annoying sometimes
3 - non deterministic, lazy
What is p1, p2?
This read like a conversation you were having in your head -- unfiltered.
Very commonly used acronyms in business/SWE contexts.
The lower the number the higher the priority, e.g. if production is broken someone might say "This is p -1" (higher than your highest priority).
P0 is "do this immediately", P1 is very important, P2/P3 are like "nice to haves." P100 would mean something like "we will never do this"
Same understanding as you.
At 3am,
P0 = get people out of bed and fix this now.
P1 = we can fix this in the morning.
P2 = we can fix this before CoB today.
P3 = push it into the current sprint.
P4 = put it in backlog.
priority 1,2 etc, more like severity of the cases
It’s a very common abbreviation; i had no issue understanding GPs post tbh
Was that the old DS 4.1 Flash or 4.1 Flash 1017?
I tried it out today, was the new one: 1017.
What reasoning effort?
Flagship gpt models are crazy efficient. 6.1 Sol is like a savant that looks at a problem and spits out the answer seemingly without thinking.
Opposite of my experience - openai models, and codex are clearly the slowest (out of 5 different models/harnesses I use on a regular basis).
Hilarious sentence: "DeepSeek-V4.1-Flash-1017 is a minor update to DeepSeek-V4.1-Flash, delivering substantial improvements across reasoning, coding, agentic tasks, and visual benchmarks." (emphasis mine)
Kinda shows their team is more engineers than marketers.
But with that version number, it's not a "minor" update! It's a patchlevel update! -Engineer
An old-school (pre-LLMs?) engineer would recognize that DeepSeek is in no way doing semver, so your major/minor/patch glossary might not apply here at all :)
Strictly speaking, it just shows they aren't marketers. Bad copy doesn't tell us anything about their engineering skill.
It's not a bad copy, or at least that's not what the person you responded to is insinuating.
Deepseep-v4.1-Flash-1017 is a minor update. A major update is something like Deepseek-v5-Flash.
> It's not a bad copy, or at least that's not what the person you responded to is insinuating.
FWIW, GP didn't write "a bad copy" but "bad copy", referring to the marketing sentence quoted by OP, rather than the model that was just released (as I presume "a bad copy" sounds, maybe like it's called that because it's [again, presumably] distilled?). "Copy", especially in a marketing context, is a written message.
Strictly speaking, it's the substantial improvements that tell us about their engineering skill.
Strictly speaking, they didn't say anything about their engineering skill
To me this implies they continued the training run for 4.1 and it improved. A 4.2 would imply they changed something about how it’s trained/how inference works, and a 5.0 would be a major update (who knows what would have changed).
This is just my headcanon though.
It's not really contradictory, minor changes can have major effects.
So they meant a "small" update - "conservative", "inexpensive"...
An example when using a less frequent term actually reduces specificity.
Changing TREE(2) to TREE(3) is a small change on the input, but a... lets just call it big... change in the output.
No, they used minor correctly.
That's one small step for a man, one giant leap for mankind.
Hoping it's minor relative to what's coming
Minor updates can provide substantial improvements sometimes, it really depends on the specifics.
Remember when we were dumbfounded why our service is slow despite having implemented quite robust caching mechanism. Turns out simply the flag that enables it was hardcoded to false. So it was minor update (changed flag to true), but provided substantial speed up.
They changed how it works substantially. I think it should be 4.2.
Quiet confidence backed by provenance :).
It's a bit refreshing to read updates from an AI lab without bullshit doomerism and overwhelming sense of self-importance from shit like OpenAI and Claude.
Also, DS is an excellent model.
I'm at 1 trying to get to 10. I move to 2. Minor update. 100% improvement towards goal.
These are not mutually exclusive.
That is how exponential functions work. Improve every aspect of a system with 1% and you may get easily 50% more performance.
It's about playing to two different audiences.
The CCP requires extensive certification for new major releases, to make sure a model doesn't do anything they don't want it to.
To get around this, Chinese labs only bump versions when they have to and usually claim it's a minor update.
Then on the other side, they're obviously trying to market to the tech crowd...
Source?
Everything in China is strictly regulated but loosely enforced. Just take that as a grain of salt
Drum Tower podcast (from the Economist, paywalled) about Chinese AI.
Would highly recommend it, it's IMO the only serious Western reporting about China.
> the only serious Western reporting about China.
as good as Gordon Chang.
Its a great update, I ran a benchmark comparing verbosity and intelligence compared to old weights.
- Reasoning effort = 1 is now actually usable and produces 2.5x less tokens than old weights, without intelligence lost.
- Overall 20% less tokens at each reasoning effort.
- Same intelligence and same tool-calling pattern.
More details in blog post https://sference.com/blog/deepseek-v41f-1017-benchmark
Improved on exactly the items that I hear people complain about with the previous checkpoint!
I've been using DS v4.1 Flash daily for nearly a month now on Agent Swarm experiments. I've tested it against GLM-5.3-flash and Opus 5.5. DS 4.1 comes out tops for me. Opus 5.5 is still the smartest, but DS 4.1's speed and price more than make up for the "intelligence" gap. I've ran Chief/Builder/Reviewer swarms autonomously for multiple days. I think my new workflow is going to have DS 4.1 Flash do all the engineering and just have Opus review its work. 4.1 Flash really is a very impressive model. Kudos to the Deepseek team.
Ok but this is a new model with significant architectural differences and new training run.
Then I guess all that it changes is that my proposed setup gets even better?
Huh? This is a new checkpoint. It has identical architecture, only different weights, and most likely just more post-training.
Much like Qwen 3.8 Next; which is qwen4-exp in a lot of the code/patches, there's usually a tick-tock cycle with LLMs. That said, it's hard to know for closed frontier models.
It's not identical, it has vision built in and asymmetric active from encode to decode.
DeepSeek delivery deliberately uses confusing version numbers.
4.1 flash was already multimodal. What do you mean by "It's not identical, it has vision built in..."
I really like the having DeepSeek Flash + Opus 5.5 as advisor/reviewer at the moment. I run through a $200 Claude Code subscription in less than 24 hours now just for a single project, and I hate fiddling with different accounts every day. Now that I get a 200 USD api credits inlcuded with the subscriptions I have rewritten my coding agent workflow completely.
I'm baffled, my $20 is enough for me to do some consulting, personal projects and side work, and still cannot outran my quotas, I'm building different kind of things: flutter apps, dotnet backend, database, front end, architecture diagrams, code exploration, tests, PowerShell scripts.
How people spend more than 200 usd monthly is beyond my understanding, are you guys so inefficient with the usage? or I am the opposite?
Without any context for what you're working on, I'd have to assume the reason you don't hit the $20 subscription limit is that you're either not using it very often, or you're working on some relatively basic CRUD app that doesn't require it to think very deeply.
I tear through my $200 subscription because I often have 3-4 agents running simultaneously, exploring performance optimizations, building features, monitoring production rollouts, or investigating production problems/customer issues in parallel.
The downside is that I do spend most of my time reading the agent's output these days.
> I run through a $200 Claude Code subscription in less than 24 hours now just for a single project
How?
Its easy when you spawn a couple of subagents in a homelab with a rack with 5+ servers.
This is how one of my claude code session looks like right now, one prompt from yesterday:
Not sure if this will help you but it helped me. I was using subagents and workflows for lots of long running tasks, and part of that work included running tests. If tests took longer than 5 minutes, the cache expired, and the whole thing was then being processed as uncached input.
I created a skill to better manage agents, including a way to keep the caches valid, and it has decreased my usage considerably. I now struggle to use my 100% each week and I hit claude hard. I have a reset expiring in 5 days and I'm trying to burn down the tokens (68% with 3d 8h left) quickly so I can use my reset before it expires.
Token efficiency just means understanding the way the system works. Best person to help with that... my old friend claude.
After I stopped letting Claude run tests, my token usage went from "gone by Wednesday" to "what can I automate this Saturday so my tokens won't go to waste".
Yeah, I should see if I can create a similar skill. The tests do spend some time to complete.
Claude: Confirmed: 195M of the 212M cache tokens (92%) come from full re-writes after 5+ minute gaps, accounting for 77% of relative cost. Now I'll check the existing wrappers so the skill builds on them.
Yesterday in my CC there was an update that added a feature to auto compact immediately prior to the cache being invalidated.
Some to think about with both Claude and skills.
Can you invert things. What’s deterministic and what needs “thinking”?
Can you create a script that runs things chaining together the deterministic parts with calling out to Claude for glue that needs “thinking”.
Don't suppose you would be happy to share the skill? Or is it fairly unique to your work flow?
I think the tests problem is really having an impact for me. Hadn't considered it before so thanks for flagging.
How do you keep the caches valid?
> including a way to keep the caches valid
Maybe doing bogus small requests to prevent the cache from expiring?
I mean in the API you can set the cache expiration time right? But I guess claude code can't do this
> my old friend claude.
Awfully new old friend.
I have some fairly large personal software projects and I still don't understand what useful work people are getting out of burning all these tokens on unattended agents.
What exactly are they doing for you that's giving you $200 of value in such a short timeframe?
As a dev who hasn't used LLMs as an embedded part of their work, and uncertain about it all, i'd be curious to know:
- How long do you reckon it would take a human to do the task?
- What percentage of features/ feature-extensions you add would have been done at all if you didn't have agentic assistance? Is there a tension between llm productivity gains and 'because we can' feature additions?
- Are many of the tasks for throwaway v1 features, exploratory tries?
- How speced out is the work you give it, eg did this task have pages of specs, how long did the spec take to write?
- How deeply will you manually review the work and how much time will it take? Will you read all the code / review at a high level and interrogate with llm / have the llm review and self iterate?
For what I am currently doing, years, if not a decade or maybe even two. Claude has already been running for 3 months constantly working on this project. I've been trying to do this several times before, but the scope outgrew my available time/life, too big for one person. But now I can actually build a production release by just setting the direction and let Claude to all the wiring, plumbing, tests etc.
I do now add all the major features I want, and write a task list where I add a few more minor tasks each week that just come to my mind. I need to write them down before I forget. I often ask Claude to implement several different implementations/algorithms and compare them. There are plenty of trade-offs, which usually end up in building configuration profiles, so the application can work optimally in multiple environments.
The specification is actually really large, and has grown over time as I've been learning and new ideas has come to my mind. I do not review/audit the code, it's massive would take me months, no chance! I do however use code inspection/audit tools like TSAN, ASAN and more. I will also spend time on full code reviews with the coming Fable 5.5 when a release is about ready. I do however review behavior together with Claude, and this is also where the specification grow, because this and that should be possible, but not this and that. User stories also helps.
Keep doing what you're doing, this guy clearly is just revving his engine because he likes the way it sounds and can afford the gasoline. I wouldnt be asking people like this for advise, they're completely off the rails.
There's no reason to blow through tokens this way, there's a 99% chance nobody uses the output (except maybe themselves).
Yes, but what was the task that they were working on? This tells nothing.
That’s the “how” we’re wondering about.
You're doing this wrong.
I swear I have the same question in my head always when I read such statements! I have never once reached 90% that includes everyday usage across multiple repos, artefact generation, device debugging, long running harness work!
I am not saying OP is not telling the truth. I just really want to understand!
I decided to try something a little weird research wise the other day analyzing about a thousand older research papers and burned through my $100 chatgpt sub + 3 resets in three hours and only finished half of them. It's the first time I've done that, though.
It’s because they don’t know what they’re doing. It’s like when openAI was using 10,000 agent clusters to solve problems… then realised the model could do the same work on its own and solve the same problem.
Anyone who tells you they have a “team of subagents” and “code reviewer” agents and “project manager” agents doesn’t know what they are doing. That’s how they burn thousands of dollars in a day.
Do you have a source for OpenAI discovering they can solve Navier Stokes with a single agent instead of a 10k swarm?
I think I’ll trust the experience of the people in the labs who have unlimited token budgets and are running experiments continuously over some anonymous rando slinging mud on the internet.
Also my own experience, which says having code reviewer agents is one billion percent useful. That comment alone completely disqualifies you as a reliable source of info.
in my experience there's a huge difference in token usage when creating new projects versus working on existing ones
Not OP, but I used to run out in a few days, for me it was a matter of suboptimal code review practices with many subagents looking at different aspects, often expanded to look at more than the diff.
skill issues, brute forcing with AI.
not reviewing its ongoing work and spawning subagents that end up looping each other
Loading tons of MCPs and never compacting
I can drain a $200 OAI plan in 2 hours flat working on my 10 projects simultaneously.
Maybe you're not being ambitious enough with these things? They are at their most fun when you spec out sprawling architecture documents and watch 1+8 astras buzzsaw through it.
I'm only at 260B for the year, though. I'm sure someone who has used more will be along shortly to chime in
E: The other subthread talking about tests is correct. Expensive models should not waste a single second waiting on tests or validation. If you cannot assign a single cheap model to be the test bitch, then just tell them to frontload everything and clean up later as a single wave.
Not meant as a bad faith response but have you shipped anything? Do you have income from this? Why 10 projects apart from fueling ADHD and "never finishing a project"?
Not OP and while I’m not draining my accounts in a single day/hour, I’m hitting limits with a $200-Claude and $100-OpenAI account.
> Have you shipping anything?
Yes, hundreds of PRs and shipped new features to 3 different platforms (day job and 2 side-business apps/platforms I’ve been developing for many years, pre-LLM). Those new features have gone out to customers and been used.
> Do you have income from this?
Well my day job for one, and I’ve been able to more easily sell to new clients with my side business. Without the changes I’ve shipped I would have been able to land a couple of the clients I’ve signed on recently.
I don’t think OP’s message implied they are never finishing a project, maybe they edited it?
In addition to the 3 income-generating projects, I have a number of other “just for me” projects. Some are mature now and only need occasional development when I want a new feature. Some are in flux as I build them out (like my “software factory”).
At this point I’m limited by compute and tokens more than anything else.
Some of us want to build and don't care that it doesn't monetise.
Would love to see these 10 projects that are being worked on simultaneously
After opus 5.5 release I find it very hard to run out on a 100 sub...
> I run through a $200 Claude Code subscription in less than 24 hours now just for a single project
I would be worried about your methodology when the subscription party ends. Make no mistake, one day, the party is ending and the people who are dining out on approaches that burn massive tokens are in for a rude shock.
I'll bite. Wtf are you working on that takes $200 a day and you're willing to pay for?
By not using existing tools and one shooting everything through the API - I.e. instead of decompile via Ghidra they will try to get source via capstone genersted assembly.
It is like putting a jet engine on a shipping container and wondering why it needs to be constantly kept in afterburner to at least scrape on the road while smarter users put wheels on the shipping container.
Interesting, coming from a similar situation, I'm interested what workflow you landed on. Opus planning + available to be called for deep thinking + opus reviwer? or?
Yeah the Claude/anthropic usage is quite arbitrary, I also run through my entire 20x pro plan in 1-2 days. Codex has slightly better usage levels, but is annoying to use for many reasons. DeepSeek flash (direct or via opencode go) is the best value model I've found so far.
This is a new model.
Mario one-shotted by deepseek-v4.1-flsh-1017 (in high mode, not even max) with sound and programmatically generated assets in under 5 minutes: https://mario-dsv4-1-flash-1017.pagey.site/
Honestly given how good the open weight models are getting I don’t see a viable business model for OpenAI and Anthropic anymore. The profit is in the compute, not the model. OpenAI and Anthropic spent all their money renting compute from others so yes they can sell lightweight model inference but they have a massive structural disadvantage now as others can always sell good-enough inference much cheaper.
There’s also zero stickiness to models and most companies now have their own internal routing utilities so devs and users can just keep the same API endpoint and move between models with a single line of code swap.
Unless something changes with the big labs and fast, they look pretty f*cked.
People will pay for more intelligent models. Astra and Fable are more intelligent than these open weight models and can solve problems that are much more difficult, and that's extremely valuable. It's not any more complicated than that.
> It's not any more complicated than that.
It is more complicated than that when the valuations and business plans are contingent on those companies AI having complete dominance at every level.
If companies only use frontier models for occasional difficult tasks, not only does that affect volume, it basically elminates stickiness and incentivizes companies to instead of using their products continuosly, find ways to move from the expensive stuff to the cheaper stuff.
Yes but that’s a tiny portion of AI use cases. Maybe 1-2% max. The big labs need a lot more than that to survive.
You’re assuming other model creators can’t build these mega models. Evidence is they’ve simply decided it’s a fools game to do so and they’re focused on where the use cases are.
Fable is behind Opus (which is also much cheaper) at this point in benchmarking.
I don’t know what I’m talking about but it seems to me OpenAI is heavily emphasizing productization of LLMs. Dots, ChatGPT Work, scheduled tasks, in-app projects etc.
Being able to directly distill their huge frontier models into Luna/Haiku-level models puts them a step ahead still.
i hate dario's politics, but i still find claude's eco system easier to use. Design/Artifacts is very convenient for people like me who has zero UI/UX expertise, and claude code's remote session works seamlessly on ios with a well-design client interface while you have to deal with ssh and typing in cli with your iphone keyboard if you are using opencode/omp etc
And yet, nearly all of enterprise money spent on AI tokens in the US is going to either OpenAI or Anthropic.
That trajectory appears to be changing. Virtually every company I see is shifting more and more to open weight models. Anthropic’s slowing growth rate is already spooking investors and that shift is just getting started.
That's why they are constantly fearmongering and trying to get other models outlawed. If nothing else, under the pretense of national security.
It passed my threshold for not caring to try new models back with DS4, I use it for literally everything. Announcements of new models in the series always feel curious, because it doesn't really feel like there's much left to fix that would improve on what's already there for any task I know of.
It's a bit like obsessing over patch releases of some software after a major upgrade already solved the original problem long ago.
I have quite the opposite experience: I can do any concrete coding change with any model, but the intelligence limit just doesn't go through more complex debugging examples.
And then when I try a smart model, like Astra medium and it solves it (I had a firmware bug in my new laptop where I couldn't charge through USB-C) on the first try (altough with a repo already having lots of information about the hardware/firmware/debugging programs/traces), I realize how much time I'm losing on the few examples that the less intelligent models are just not able to debug and solve.
Also they can work in completely new domains, like Astra and Opus 5.5 starting to be able to use something like Blender. Currently they still often do horrible topology and basic UV mistakes, and the animations are really bad, but who knows where they’ll be in 5 years.
I don’t think these are completely new domains in training. Like I’ve taken up offers to help code up AI training material just to see what they’re up to and get a sneak peek at how the sausage is made, no idea who these front companies are working for of course, but they pay for all kinds of fringe cases and Blender wouldn’t even be considered fringe in what I saw at the beginning of the year. (I was only there to observe, never submitted anything, so I didn’t betray you guys by helping train your replacements.)
5 years? Look at the leaps since last November. Since the Spring, even.
Very cool!
Is your laptop something like framework which has open source firmware?
I(and many other people) have a similar issue with usb-c charging not working consistently. I've been meaning to try one of these new smart models to help me debug the issue.
ASUS Zephyrus, but it was able to reverse engineer the ports and I was able to download the firware from the controller itself, write an emulator that emulates running the firmware on multiple controllers with different instruction sets (the USB Power Delivery controller is RISC-V) and the communication between them. Also it found out how to read the communication buffer that was on the devices for debugging, which helped it understand the USB data packets that were doing the handshaking and voltage/current negotiation.
I had to reboot my laptop a few times though until I got to this, but it was a fun experience anyways.
At this point open source isn't as much advantage as long as the firmware is not encrypted :)
Wow, that is impressive. I've been meaning to do this by hand to fix the issue on my Lenovo IdeaPad for a while but haven't gotten around to it.
I'll try handing it off to one of these new frontier models.
Yeah, but for some reason people think they need SOTA to help them work on their side/indie projects.
Like if we needed PhD intelligence for productivity tools and SaaS platforms, why weren't we hiring PhDs to build things before.
99.99% of people do not need SOTA ai.
Speed and price. You should try the 4.1 Flash on one of the fast providers. It's pretty crazy how fast this model works, and it's great for iterative software development because you can steer it fast to where you want to go.
https://www.lithosai.com/
https://www.coralbricks.ai/
How are these high-speed providers close to matching the API rates from deepseek?
Seems that V4.1 Flash is quite cheap to serve. Especially how it implements the cache, meaning the cache pricing can be very low depending on how you implement it in your inference engine.
Depends on your use case. But if you do software development with a harness such as oh-my-pi, your cache rates are probably somewhere above 98%. Coralbricks has free cache reads, meaning it's much cheaper than the Deepseek API rates.
Lithosai is more expensive especially in the ultra tier. It's if you want extremely snappy and fast responses, which is important e.g. for a personal assistant where every second counts.
If you generate a lot of output, you're better off with the Deepseek API rates.
Ive never needed an assistant pre or post ai to reply to me in 100ms at a clip of 500 words per sec. See you in 15 min, Im making some cofeee.
Where does one even get decent ZDR provider? I'm wary of sending sensitive data but not at a scale where I would build my own machine.
You just search around, or heck, just ask Google for inference providers that are ZDR and whatever else you want. Just make sure to read their terms/privacy and review the security center. You will have to request access if you are wanting to get into the deeper details.
Outside of major players like Amazon/Google which have compatible offerings available, I believe together.ai, fireworks.ai, and deepinfra.com are, as best as I can tell, US-based (if that matters in your situation) SOC2/HIPPA/ZDR. Each of them offer serverless model hosting with caching, which is going to be the cheapest overall. Most of them also offer private AI deployment which is more expensive but your entirely own instance. Some of them more heavily quantize their models (deepinfra) so take that into your workload considerations. I included deepinfra because they tend to have some the new models sooner than others or have models the others don’t (like mimo). IMO, together > fireworks > deepinfra.
However, you’ll have to keep in mind that for true ZDR you give up caching which increases your inference costs, so it depends on how “Zero” you need the zero to be and on what timeframe. If you just don’t want your data being trained on, normal use is fine imo. If you are working on something actually sensitive (like, because of NDA agreements and such), then they might not be the right choice and you may be looking at more expensive offerings or building a machine.
One note: you have to pay attention to the source of the inference. Many (similar to open router) offer models that are pass-throughs to a providers API, but they are pretty clearly marked. Just make sure the model is marked as serverless when evaluating new ones. This is where a proxy comes in handy so you can prevent your LLM from accessing a different models from the specific ones you’ve vetted.
I like privatemode.ai for better than ZDR, encrypted attested hardware (in EU)
Scaleway have DS4. I don't think they call it ZDR, but their framework is very strong https://www.scaleway.com/en/docs/generative-apis/reference-c...
Same. I'm just happy sticking with them and am consistently impressed. Cost is so good and kind of keeps you honest.
definitely could use less thinking time and lower token usage, which is what this bump is focused on
> improve on what's already there for any task I know of
What about "good governance"?
Does it have zero hallucinations? Does it reflect thoroughly on every single idea it uses, verifying its solidity?
A negative datapoint - We have been evaluated prior Deepseek Flash models (both the original 4.1 flash and the smaller 4.0 flash) and they have been overthinkers. Our workload is well defined but moderately difficult coding tasks and among open weight models GLM5.3 flash performs best with Qwen3.8 flash next a close second. DeepSeek 4.1 flash fails at lower thinking levels and spends a huge amount of tokens (an order of magnitude more) at the highest thinking levels where it is able to successfully complete the tasks. I will evaluate this newest release since they specifically claim "the model reduces task execution time and output length", but keeping my hopes low.
v4 was a worse overthinker. 4.1 mostly fixed that. It still prefers to plan everything before making a single file change.
If I'm impatient I switch to Luna for a single turn just to get it started and let DS continue.
What is your workload?
I have started phasing into DeepSeek v4 (Dwarfstar quant from Antirez) as my daily driver and it's going very well. It can one shot iOS apps on my Mac Studio, it's a little slow and sometimes needs retries but can definitely do medium difficulty real work for me.
I didn't try the last flash model because I would need to use SSD streaming (I have 128 GB), but the benches for this release are so compelling (and it using less tokens to get tasks done) it might be worth the throughput hit. I will probably give it a try when I have some downtime.
It feels very freeing to be doing useful work on a local model on my own hardware with my own coding harness. The TCO argument is starting to look compelling too.
Try the Qwen 3.8 flash next model on DwarfStar - you will be very surprised. I’ve been using the 4 bit quant almost exclusively after having DS v4 flash be my daily driver. Qwen is faster and seems to be much better at coding tasks.
This is a new model, not the same as a different one.
Above GLM 5.3, nice! That is a very capable model. I do wish that we will see some smaller "flash" models again though - running this one locally is tough (very costly).
For comparison, close to Opus 5. Far behind Sol 6.1 and Opus 5.5.
Probably a pretty decent model for 90% of the usecases!
I've been doing a lot of hobby work with GLM 5.3 Flash and I can count the amount of times that I need a stronger model on two hands.
I have access to unlimited tokens at works and Opus is great but also often overkill if you know exactly what you need.
I use Opus for planning, through API on medium thinking, it doesn't cost anything significant and the rest can be happily done with deepseek or glm flash. Using sol 6.1 for review loops at the end. All running on pi.
A very complex task tends to cost like 10$ max. It's comparable to a subscription with enormous flexibility.
It's a steal.
Honestly, I don't know what people see in Sol 6.1. Opus 5.5 is so much more inteligent, as in proactive, avoiding stupid mistakes, where it contradicts itself, ignores the context. Sol 6.1 has a gold-fish memory, it's as annoying for me to work with as Opus 5 was.
I'd rank Sol 6.1 only slightly behind Opus 5.5 in intelligence and both significantly ahead of anything else I've used (though cheaper/faster models are totally fine for a lot of tasks). My only real annoyance is 6.1 Sol is one of the slowest models I've used, which does make Opus 5.5 more pleasant to use.
You can override the context length of Sol to 1M
Without this change I hated Sol and was burning tokens way faster due to all the forced compacting.
Since making this (obscure/hidden) change I enjoy Codex much more.
That's not actually the problem. It forgets the instructions from 15 minutes ago, without context compression involved in-between.
By next year we will likely see GLM 5.3 level performance in Qwen 4 MoE models that can run on most higher end consumer systems.
I don't think there is a reason to care about any of it, just get a few dollars on a few openrouter or similar platforms and just use the models for now, wait for cheaper models that are good enough for you to become available for your hardware.
Sure you might never see frontier performance on consumer hardware but do you really need that though? Does anyone? And if you ever do just get a 30-day sub for that work, and cancel it.
Yeah performance at a given compute budget will likely continue to improve.
We are running Qwen 3.8 27B FP8 with 4x 5060 16gb at the office for a small development team. It is doing pretty well for implementing backend and frontend tasks, and is already useful. In 6 and 12 months from how I expect it to get even better. Some planning is still done with a stronger model, but even that is maybe not strictly necessary.
Qwen 3.8 wasn't that nice in my experience Qwen series has lagged a lot but I think Qwen 4 would be better they cleaned up the architecture.
And we are likely to see many companies start building 27-32B MoE models custom built for use cases now that Rubin is going to make training costs much lower.
I expect training costs to get under 100k USD within a year for small models with another 100k maybe for RL.
I think just having all your own data from claude and stuff on your system that you can feed these models and have companies offer great base 27B models will be how it works.
Cant you run it on a pair of DGX spark? Sure it's expensive now, but at the initial retail price that was what, 8k in hardware? Not exactly unachievable
No, this is not a small model. I believe 4x Sparks is possible but 4.1 is pretty big for a flash model.
MiMo 2.6 Flash is the most capable model you can currently run on 2x Sparks without resorting to a quantised version of something.
I meant quantized. You can get an EXL3 quantized version thats almost the same performance.
My experience running quantised models is that they look the same at first but are all noticeably worse once the context size gets beyond anything trivial.
Still need 2 or 3 days to properly quant this. But it is really good model to run if you have 64GiB or larger RAM on Mac. On M5 Ultra, you get ~3400 tok/s for prompt processing and ~50 tok/s for token generation, quite usable state.
Forgive the ignorance, but how are you running a 763B parameter model on 64GB of RAM?
SSD weights streaming. The actual weights is 552B, then engram.
which engine?
https://releases.drawthings.ai/p/public-beta-of-local-code-b...
By “quantitizing” it
The smallest quant is still over 100GB, so I don't know how that fits in 64GB RAM. It would have to be partly streamed from disk.
Here are some pelicans. The "max" one took over 11 minutes: https://tools.simonwillison.net/markdown-svg-renderer?url=ht...
Definitely an improvement on deepseek-v4-flash-0731: https://tools.simonwillison.net/markdown-svg-renderer?url=ht...
You should start asking for animated three.js pelicans too, the 2D pelicans seem to have saturated.
here you go! Results are trickling in https://fenderog.com/art-crit/page.html?p=surfing-pelican
Faithful to the original theme. Apropos. +1 on adding the prompt as long as it’s affordable and easy enough Simon, and not too distracting/noisy/time consuming all considered!
This is in "high" mode .. 40 seconds only: https://ibb.co/5XTRKNJv
Looks much better.
Wow. Is DS the only model that consistently puts one leg behind the frame? I don’t even remember Opus ever doing that.
Some of the most recent Claudes and GPTs have done that too, and Gemini has done it for a while.
The best bicycle is "low".
I essentially only use DeepSeek. Almost always the flash variant. I've found that I don't really need frontier models if I'm doing most of the thinking myself and just using the agent to output code quickly.
Getting a lot of errors from OpenRouter's endpoint "Failed to get response from the AI model; retried 5 times (total retry wait time: 60.71 seconds) Last error: 429 Provider return error"
It works when it works, but will probably be flaky AF for a few days.
Works well with Copilot:
700000 because there's no detail yet on the safe completion limit for most harnesses (OR lists 1M for both which makes no sense). COPILOT_OFFLINE enables all calls except to OR (no GH servers).I haven't put this into test as much yet.
As a Luna6 proponent, this has a potential as it looks slightly more intelligent and at the similar price point.
I can't wait for a smaller moe dinstillation from this excellent model
Thank you deepseek, we are eating good today.
I am a huge DeepSeek fan, but it cannot come close to Opus 5.5 on UI design. A $20 Claude subscription that I use some months and cancel the others along with Deepseek API is great.
Once read, I think from G Huntley, "pick a model and stick with it," they have been making that very easy, daily driver since V4.
Looks very impressive especially considering the price and speed. Let's wait for AA to confirm how well their benchmarks hold up.
You trust AA since the sol fiasco a month or two ago?
What did I miss?
They updated their methodology when the new openai model didn't rate higher than the old one, or something like that.
I do t think it was a fiasco. I thought the updates were good. They should do more work on benchmarks. There has been good work on testing benchmarks for example to ensure they are accurate, and as issues arise they should change the benchmarks. The point of AA and Arena is trying to understand how to think about models.
Why not just try it yourself?
It costs money to run serious tests. We're way past the time when you could get an idea of the new model in a couple of tricky prompts. Now I need to compare a whole resulting project to get a clear view. If someone else proves the benchmarks don't hold up in reality, then it saves us further tests.
At this point I am way more excited by DeepSeek releases than anything from OpenAI or Anthropic.
The speed and cost/intelligence ratio is just outrageous.
The native web search api is 10x cheaper than OpenAIs. I hope they will release the web search api as a standalone API.
I may be missing something so apologies if I'm telling you something you already know but: there are plenty of web search APIs for agents - the Brave one is the one I use: https://brave.com/search/api/
Tavily one is free, too. Easy to use cli with a skill.
I gave it a kick and got mixed results. For example, searching for a Microsoft product couldn't find its landing page, and searching for a niche open source package landed in the Debian repos for the top hits, but the home page was in the first 10 hits.
Lovely to see they're doing their own search though, I love how many folk are actively working on search clones at the moment.
Already released https://api-docs.deepseek.com/guides/search/
Have security companies considered that these open models could be used as sleeper agents even when they're deployed outside of China? Is that a serious attack vector or more like science fiction?
There are theoretical attacks involving malicious weights. But the risk profile is so outrageously in favor of them not doing it that you'd essentially have to be an anti-China conspiracy theorist to find it plausible, let alone likely.
Those are crazy numbers for a Flash model. Hopefully they can scale what they are doing to 4.1 Pro.
By far my favorite model.
At the beginning of the month I realized I had more tokens and more capabilities on Opencode at 8$/month via DS 4.1 flash, than I had on a 200$ account in June with the american providers.
Literally both more tokens and capabilities than Opus 4.8/GPT 5.5.
Even if you pay per token, you're still unlikely to outspend american providers.
I've been telling people this for a couple of months now, I'm actually struggling to use up all my tokens.
But this is a new, different model that just came out.
What does it matter if you’re not hitting your cap with OpenAI/Anthropic in the first place? This thread is apparently populated with the highest of powerusers running multiple OpenClaws. Us peasants aren’t concerned with meters or cost/token - just quality - and nobody comes close to OpenAI or Anthropic in that regard.
Pretty funny to say "us peasants aren't concerned with ... cost/token", as if most people ("peasants") won't find it absolutely insane to be paying $200/mo for a subscription.
Peasants don't pay for the subscription. The employer does.
How could that be considered insane for a professional tool?
In southern Europe (Italy, Greece) that's well beyond 10% of the average pay.
And southern Europe isn't exactly the poorest region around.
This is a great model. I know we are all looking for the next best thing, but I'd be pretty happy running this for awhile.
I use Opus 5.5 and swarms of about 50 deepseek models to reverse engineer old videos games. The scale is roughly 600k loc generated from pure asm, simply by knowing what the old compiler is. Truly staggering stuff.
Registered just to ask you.... 50 instances? Which harness? How? Why? Can you post a video demonstration?
Caveat: I get a budget of free ai tokens at work, as I work for an open source AI inference provider.
My goal is to reverse engineer an entire gamecube game -- I'm pushing an existing community decomp that was stalled to the finish line, solo, with lots of AI help.
This work was directly inspired by https://decomp.dev/ -- which has just shut down a few days ago because of, ironically... people like me who enjoy decompiling things with AI, so I'm still weighing / grappling with the ethics of how to release my work, it's murky waters at best.
I use Opus 5.5 as the orchestrator, as well pi harness because it's somewhat lightweight. But for the really hard functions I don't use deepseek I use opus managing other opus agents.
I'm working to release something soon as proof that I can do a 100% decompilation job, but the main idea is that when decompiling a gamecube game your sources of truths are individual functions or files, so you can build off that primitive.
Each ai gets a worktree, and a shared giant context file that's basically a record of every experiment that has happened. They're assigned tasks to essentially do as much as they possible can for each individual funciton. After about an hour claude uses a script to merge the work that's actually correct (verified by tooling that checks if the c goes to asm) and then there are usually 3 rounds of reviews by independent ais, because there will always be some slight mistakes. The techniques used as well as the mistakes the models make are recorded in the giant ledger. There's a bunch of random one off scripts such as stuff to clean up, and tools to make sure the agents don't accidentally google stuff / delete stuff / use too much space.
Claude steers the ship via ssh on my strong linux tower at home so it keeps my laptop nice and cool and at this point I just check in from time to time, I am basically a manager, and Claude has it's own workers.
Last point -- I don't know any c besides the basics, I've only been "coding" python / ts for a couple of years, but even then with the extreme strength of models the future is managing agentic swarms, so I'm leaning in as much as I can. I really see this as the future of large scale projects. For some people it's a mockery of decompilation, but, my goal is the finish line, not the journey.
Could you please share your methodology?
See here https://news.ycombinator.com/item?id=50135410
What would be the best local harness to run this? It feels there are 1000s now.
Exactly, there is no ”best.”
Here's a small benchmark that compares the new checkout with the previous one:
https://slopgrid.com/deepseek-41-horses
Another big win for opensource AI.
BTW who would have thought China would be the champion of freedom and opensource in the AI arena, while US is fighting for corporate/government control.
Well, at least mostly everyone outside the US.
Anyone that has observed China since the 90s.
There are different types of freedom and China's bet is that as long as they provide enough economic freedom the public will be ok with somewhat restricted social and political one.
Yeah, I had a somewhat odd conversation with a Chinese acquaintance last year when she got to talking about how America sees China. Her point was that, "How am I not free? I can do whatever I want, go where I want, buy what I want. I can talk to you and say whatever I want; nobody is here listening to me."
The fact that she can't call Xi "Winnie the Pooh" on Douyin wasn't something that occurred to her, or maybe that she was aware of at all. I didn't ask. I don't like to get into these discussions when I'm in China.
There is no such thing as “free speech” anywhere. It’s all a spectrum and the Americans in particular are indoctrinated. It’s a red herring to distract from the real social issues. People in America have some of the worst healthcare in the world. But hey at least we have “free speech”.
Fixing healthcare or economic policies is typically an act of congress type of thing - not even an amendment. Large popular support - and it can be fixed. Fixing freedom of speech or the ability to vote at all - once it is gone - usually takes a price payed in blood. Sometimes rivers of it.
I can't call AI "AI" without being termed an "enemy" by the President.
That economic freedom is a fantasy fabricated so that the king can rob people fortune created by the themselves in the name of greater good.
That's an image China wants the rest of the world to believe. Towards inside it still fits the typical stereotype.
It's better than GLM 5.3 and Kimi K3 and way cheaper.
that's honestly insane. it's like the go-to for when you don't want to use opus 5.5/sol 6.1
The go-to is a local model.
$$$
Limp-to, maybe.
With Strata, running GLM 5.3 is possible.
The new model is amazing. Had it build a full Angular Material & CDK components showcase via DeepSeek Harness and it easily did the best job of any recent models I've seen.
Does DeepSeek provide the inference code along with the weights? I read here recently that for models where only the weights are released, providers have to implement which could be different than what the labs themselves use.
There's basic reference inference code released with the weights, yes.
https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash-1017/...
> providers have to implement which could be different than what the labs themselves use.
I'd imagine (most) providers just take implementation by vLLM/SGLang.
I wonder how you would be able to reliably rely on a model like that for agentic work, even for implementation only after careful planning, how are you making sure you are not wasting tokens reviewing and correcting it over and over again ?
Yeah, I have the same question. I'd like to incorporate cheap & fast models like DS into my workflow to reduce my reliance on OpenAI/Anthropic, but after trying several such models and having to have a frontier model review/fix/redo everything multiple times per task, I gave up and simply have the frontier models (Sol/Opus) do everything. All of this fits pretty comfortably in my monthly subscription.
Maybe the equation changes if I exceed my subscription usage, and need to fill the gap with something else. But it seems like a huge waste of your time to go out of your way to use a cheap model if you have the tokens to spare in your subscription.
This and (the not quite yet open) Kimi K2.8 are my main coding engines, and I find them incredibly cost efficient. I find Deepseek in particular is really good with Svelte work.
Right now I use each of them in their own respective coding harnesses, deepseek-harness and kimi-code, but I do wonder what they are like inside something like opencode.
Very keen to try this out tonight. What coding harness do people like to use with these models?
DeepSeek Harness works very well, also used Codewhale which is not too shabby
You can stick with Claude Code if you're already using it. DeepSeek has an Anthropic-compatible endpoint, so you just need to set ANTHROPIC_BASE_URL=https://api.deepseek.com/anthropic and ANTHROPIC_AUTH_TOKEN to your DeepSeek API key.
Pi is very good with open weight models (proprietary ones too)
Pi if you're a minimalist
OpenCode if you want a good, opinionated setup
Why no Jinja-format chat template?
I'd never heard of the BabyVision benchmark they do so well on. Pretty interesting stuff:
https://arxiv.org/abs/2601.06521
Is it already updated to the newest version on Opencode?
Yes. It’s actually free to use at the moment. Of course with a quota.
Both the free (seems limited to ~200k context) and metered variants in opencodego show the release date as 1788998400000 → 2026-09-10, which implies they're not the newer checkpoint
Is it coming to OpenCode et al?
Presumably. The current opencode slug for the model is just 'deepseek-flash' suggesting they will just update that one going forward, not keeping versions in parallel.
and do we know which version runs on opencode at the moment (on the free and on the paid version), is there a way to find out?
The only way I'm aware of is to check their twitter feed. They moved to Deepseek v4.1 flash in September
sure, but which version are they running. they could be running an older revision
Just yesterday there was just one public revision of 4.1, and they listed they were using 'Deepseek v4.1 flash'. I assume they will update their website if they update the model version (and use the version in the name).
My favorite model just got better. Gotta love what deepseek is doing
they must've been using this since 2-3 days when it became free again in OpenCode, it's noticeable smarter, beats GLM
beating K3 and GLM-5.3 is kinda insane
And Christmas comes early this year.
DeepSeek v4.1 Flash was a beast.
It was already cheaper and more powerful than Haiku 5.5 which was released like a full month after it... by a trillion dollar american "frontier" company.
Can't wait to see what this one has in store.
"a minor update" they say before showing off benchmarks that display a substantial improvement.
Probably a week ago someone posted a thread with the title "Why isn't the industry freaking out about DeepSeek 4.1 Flash?"
I have been intrigued about Chinese models for a long time, but never tried anything except an early Qwen. My daily driver was Luna, with occasional elevation to Terra or Sol - all on 5.6 as I found the output of 6 jarringly incomprehensible.
I still had 95 USD on my OpenRouter account, so I thought I would give DeepSeek Flash 4.1 a try.
I think now I can understand the person who created that thread.
Here is the setup I settled on, after a lot of refactoring and calibration of the workflow:
I have a "plan" prompt which explains to the agent (Luna) that I want to have a discussion about an idea I have in mind. The agent should use the Socratic method to pull out all the necessary info from me using pointed questions. When the agent feels it has a complete picture, it should elevate to a higher level model (Sol), and ask its review. If Sol has complaints, Luna keeps asking me until the points raised by Sol are clarified. The review rounds with Sol continue until Sol says "ready to implement".
At this point Luna creates four files in a directory I specify:
1. PLAN.md: the implementation plan with all the details that came out from the design discussion 2. PHASES.md: implementation phases 3. DECISIONS.md: design decisions (this serves the same role as ADRs aka. architectural decision records) 4. DIARY.md: append-only log of what happened during implementation
Files 2-4 are written by the implementing agent. Luna only adds headers to explain the purpose of the files.
Once these files are there, I start up a new Pi session, switch to Deepseek 4.1 Flash and then point it to this plan folder.
And then it goes brrrrr... :-)
Here is what I implemented in the past week using this method:
1. A capture/analysis/decompilation pipeline for C64 demos: I ask the agent to start a selected demo through VICE, and say "on" or "off" when I want it to start/stop capture. The capture process uses VICE's builtin trace functionality: the trace log is piped into Clojure where it is saved into an optimized binary format (chunks of trace events). Once I have everything captured, I unleash my analysis pipeline which checks the write patterns of the executed code: what it wrote (VIC, SID, CIA, memory) and when (rasterline, cycle). Then it splits the execution trace into segments by type deduced from the write patterns (decruncher, loader, raster, music) and then decompiles each segment into both traditional ASM files and a s-exp based syntax which roundtrips back into the in-memory representation. Once I have the results of the analysis pipeline, I can tell the agent to analyze any demopart and it can explain how the effects are built down to the tiniest timing detail. As a first test, I asked it to create a clean-room implementation of Fairlight's legendary intro and two days later I had it, with everything in sexp assembly, including the music routine and its data. (The reason it took so long was that this was the first decomp and it provided a lot of feedback to the pipeline development.) Finally I had it write an assembler, linker and packager which creates a PRG from the assembled and linked artifact, then runs it through VICE. Done.
2. A DSP compiler which lets me compile sexp forms of a C-like DSP language into native DSP kernels running on the JVM, JS and WASM. Currently I have about 600 tests which verify that the behavior of the three targets matches that of a reference interpreter written in Clojure. I have a full numeric tower with i32, f32, f64, support for arrays and structs, typical control flow constructs, and a library of DSP building blocks like oscillators, filters, LFOs, delays and a chorus effect. By the time this is ready (end of the weekend?), I will be able to build synthesizers like Synth1, Surge XT or Vital in Clojure/Script, compile the DSP kernels in the browser, choosing between JS or WASM backends, driving the kernels via AudioWorklets and sequencing them from the host language.
There are two power tools at play here:
1. Deepseek Flash 4.1 2. Clojure and ClojureScript
Deepseek is a BEAST. It just goes on and delivers, phase after phase. What it needs: clear instructions - this is covered by the planning process quite well, but I still find it needs some steering from time to time (the right nudget of wisdom dropped into its thought stream at the right moment works wonders)
And Clojure... well. That article PG wrote about LISP (Beating the averages): what I witness now proves all of those claims.
The agent LOVES Clojure. Especially the REPL (not so much the parentheses).
Winning strategy: give it two tools: start_jvm and clojure_eval. Implement everything else it may need - tools to start up HTTP servers, compile ClojureScript, add dependencies, capture a C64 demo, run Playwright via spel, etc. - as lightweight Clojure namespaces which provide APIs to all these functionality. Let these namespaces be thin: the namespace and the functions therein shall have docstrings which tell the agent what the functions do, but the functions themselves shall be lightweight proxies to the real implementations in sub-namespaces. Then point the agent to these top-level API namespaces via your AGENTS.md. Bam, instant discoverability with minimal context usage. It will do EVERYTHING through clojure_eval and the round-trip time will be phenomenal, a perfect match to the raw speed of the agent.
...
Now all I need to figure out is what will I do when all my dreams become manifest in flesh.
Isn't there a lot of overlap between PLAN, PHASES and DECISION that pollutes the context ? Whereas a simple PLAN.MD to put the harness on track and an ARCHITECTURE as a single source of truth to prevent it go offtrack would suffice and gives it more space for higher level of reasoning ?
These files came to be after I've seen Deepseek create them for itself to track progress.
Then I turned what Deepseek did into an upfront policy.
This seems to work quite well generally: watch the agent as it works, and think what could be provided to make its life as easy as possible
Have you thoroughly tested a slightly simpler workflow with one less stage but a frontier model doing the initial planning? Luna is amazing but it's a deliberate trade off designed for speed and cost savings. You can also give them a rule about speaking clearly.
The idea behind using Luna was cost-optimization.
I wrote a Pi extension called "model-escalation" which lets the model request an automatic switch to a smarter model for the next turn if it thinks the problem it is struggling with is too complex or if the user explicitly asks it to do so.
With this extension in hand, I could split the work between Luna and Sol, resulting in a ~95%-5% split. I don't have to switch to Sol when review time comes, Luna does it automatically. Sol tells its remarks, then Luna gets the control back and continues the back&forth (if there are still complaints).
Pretty cool, thanks for sharing.
at this point there are 100s of benchmarks.
so AI labs can just pick which to show
So this is the reason for the crash a few days ago.
What crash?
Deepseek servers went unavailable for 5 hours. They compensated that by rewarding the previous week usage as credit though. Pretty nice.
Ai/tech stock crash
oh is “Flash” a cool sounding word again since Adobe abandoned it?
Performance charts and benchmarks are here:
https://www.reddit.com/r/DeepSeek/comments/1x84wke/deepseek_...
Where are the Unsloth quants for DS-4.1-flash?
Web search: $0.75 per 1,000 calls
That's way cheaper than openai and other web search providers. I have no doubt that AI demand justifies all the compute money big tech is spending now, but I also think there are so many over valued VC funded companies that are going to crash in the next few years.
- how good is it at reversing a game exe file?
Anyone using DS as CLI coding agent which harness do you recommend? Is there anything as good as CC/Codex?
I really like pi, though it does require some time in setting it up. But it has greatly reduced my token spend. The relevant Deepseek Harness https://www.deepseek.com/en/harness/ I hear is also good and popular.
I use deepseek via fireworks using Claude code.
When I type deepseek on my CLI it launches Claude with my fireworks API key. Works very well and feels almost free.
PSA Deepseek train on your session if you use their endpoint and/or harness.
I'm a communist so I'm fine, but you may not be, so be aware
PSA OpenAI&Anthropic train on your session if you use their endpoint and/or harness. I'm an American! so I'm fine, but you may not be, so be aware
So is this the first model that passed China's new export controls introduced yesterday?[1] It goes live November 15, but temporary pre-registration requirements took effect immediately after the announcement, ~20h-ish ago.
[1] https://www.reddit.com/r/LocalLLaMA/comments/1x7g3zq/china_p...
Presumably they registered, but that doesn't mean the model passed through export review. It just means that some bureaucrat now has the opportunity to check the product characteristics they reported and decide whether that kind of thing should be allowed for export or not. (Though with the weights out in the open, that's maybe a bit pointless.)
Try adding a new comment
We're in a crisis
Censored, no surprise
Oh mince, mon immense meurtre des humains congois dans le fruit
Those benchmarks can't possibly be real, right? Right? <insert padme 4-panel meme here>
At this rate the frontier labs will be Anthropic and DeepSeek, with OpenAI struggling to keep up.
OpenAI's models -- and even Google's Argon -- have particular niches where they shine. For GPT-6/6.1, it's 3D modeling; Astra is hands-down the best at making .step files, Blender files, etc. For Argon, it's natural language writing and everything that hooks up to Google, e.g. search and maps-based itinerary planning.
Anthropic and DeepSeek (+ GLM) are definitely solving coding, though!
They're all quite different tools.
Artificial Analysis recently released a dedicated 3D benchmark, where Astra lags behind both Opus and Sonnet 5.5 (and yes, open weight models seem to be quite far behind).
https://artificialanalysis.ai/articles/aa-3d
Interesting!
For me it's large files in Blender (in some cases >400MB), mostly, where Astra excels. I gave it a look, but I don't know if the benchmark encompasses that sort of thing.
.step files are simpler pieces of straightforward code, so it makes sense that Opus and Sonnet 5.5 do very well...
its unfortunate they didn't include deepseek in the initial batch of benchmarks
I believe these are the last sets of super-high-end models available to public (open or closed).. these companies probably will now start to withhold top-of-the-line models for themselves, governments, and their super-high-paying partners only.
That hasn't been the way to bet. DeepSeek and the other Chinese firms show no signs of slowing down, so if OAI and Anthropic halt releases and "pace the frontier" they'll just fall behind. (And there's a chance that high-end AI will become fully commoditized, which suits DeepSeek just fine.)
Besides, Fable 5.5 is reportedly releasing within the next few days, probably on Tuesday...
China just announced tightened export controls on AI models, for what it's worth.
News from a near future:
Here's How Distill APT Used McKinsey's Private Access To Top Anthropic Model To Help The Chinese Keep Up With The Us Top IA Labs Despite Access Restrictions
We have the weights, we have the papers, nothing preventing development on this foundation.
Except compute
idk, sol 6.1 is both cheap and capable. i've been using it exclusively, i don't need to care about usage, and it's really smart.
yes maybe opus 5.5 is stronger and flash is cheaper, but 6.1 strikes a good balance imo.
OpenAI models don’t come close to competing with deepseek in cost quality ratio. I have a $20 plus plan and I run out of my 5h limit in literal minutes, every tick of the harness is like 10% of my 5h limit.
You can in contrast get far more done with $20 worth of deepseek api spend. No comparison.
maybe, but sol 6.1 is far more capable than deepseek, and i love deepseek
i have a feeling that Sol is just Astra low, but with lowest possible priority on the compute.
I believe it's Astra quantized. It was briefly named astra-lite.