Senior software engineer with 46 years of experience (since I was 7). LLM inference hasn't been too useful for me for writing code, but it has proven very useful for explaining my coworkers' code to me.
Recently I had Gemma3-27B-it explain every Python script and library in a repo with the command:
$ find -name '*.py' -print -exec /home/ttk/bin/g3 "Explain this code in detail:\n\n`cat {}`" \; | tee explain.txt
There were a few files it couldn't figure out without other files, so I ran a second pass with those, giving it the source files it needed to understand source files that used them. Overall, pretty easy, and highly clarifying.
My shell script for wrapping llama.cpp's llama-cli and Gemma3: http://ciar.org/h/g3
That script references this grammar file which forces llama.cpp to infer only ASCII: http://ciar.org/h/ascii.gbnf
Cost: electricity
I've been meaning to check out Aider and GLM-4, but even if it's all it's cracked up to be, I expect to use it sparingly. Skills which aren't exercised are lost, and I'd like to keep my programming skills sharp.
I just go outside when my internet is down for 15 minutes a year. Or tether to my cell phone plan if the need is urgent.
I don't see the point of a local AI stack, outside of privacy or some ethical concerns (which a local stack doesn't solve anyway imo). I also *only* have 24GB of RAM on my laptop, which it sounds like isn't enough to run any of the best models. Am I missing something by not upgrading and running a high-performance LLM on my machine?
Does the cost of a new computer not get factored in? I think I would need to spend $2000+ to run a decent model locally, and even then I can only run open source models
Not to mention, running a giant model locally for hours a day is sure to shorten the lifespan of the machine…
> Not to mention, running a giant model locally for hours a day is sure to shorten the lifespan of the machine…
That is not a thing. Unless there's something wrong (badly managed thermals, an undersized PSU at the limit of its capacity, dusty unfiltered air clogging fans, aggressive overclocking), that's what your computer is built for.
Sure, over a couple of decades there's more electromigration than would otherwise have happened at idle temps. But that's pretty much it.
> I think I would need to spend $2000+ to run a decent model locally
Not really. Repurpose second hand parts and you can do it for 1/4 of that cost. It can also be a server and do other things when you aren't running models.
The computer is a general purpose tool, though. You can play games, edit video and images, and self-host a movie/TV collection with real time transcoding with the same hardware. Many people have powerful PCs for playing games and running professional creative software already.
There's no reason running a model would shorten a machine's lifespan. PSUs, CPUs, motherboards, GPUs and RAM will all be long obsolete before they wear out even under full load. At worst you might have to swap thermal paste/pads a couple of years sooner. (A tube of paste is like, ten bucks.)
Not on current hardware. I have an AI voice bot running 24/7 on a Mac Mini in my office (provides services for a dedicated server for a video game) and the amount of power used above idle is minimal.
IMO you're better off investing in tooling that works with or without LLMs:
- extremely clean, succinct code
- autogenerated interfaces from openAPI spec
- exhaustive e2e testing
Once that is set up, you can treat your agents like (sleep-deprived) junior devs.
Been using gRPC with json transcoding to REST on a greenfield project. All auto generated clients across 3 languages. Added frontend wrapper to pre-flight auth requests so it can dynamically display what users are allowed to do.
Claude Code has been an absolute beast when I tell it to study examples of existing APIs and create new ones, ignoring bringing any generated code into context.
Autogenerated interfaces from openAPI spec is so key - agents are extremely good at creating React code based on these interfaces (+ typescript + tests + lints.. for extra feedback loops etc..)
I use Copilot, with the occasional free query to the other services. During coding, I mostly use Claude Sonnet 3.7 or 4 in agent mode, but Gemini 2.5 Pro is a close second. ChatGPT 4o is useless except for Q&A. I see no value in paying more -- the utility rapidly diminishes, because at this point the UI surrounding the models is far less important than the models themselves, which in turn are generally less important than the size of their context windows. Even Claude is only marginally better than Gemini (at coding), and they all suck to the point that I wouldn't trust any of them without reviewing every line. Far better to just pick a tool, get comfortable with it, and not screw around too much.
I don't understand people who pay hundreds of dollars a month for multiple tools. It feels like audiophiles paying $1000 for a platinum cable connector.
Thanks for asking. I’ve been thinking about this lately as well, but I always come back to the idea that paying for online services is worth it for now, since this tech is evolving so quickly that a stable local setup wouldn’t justify the time spent tinkering — it could be outdated by tomorrow.
You can get used RTX 3090 for $750-800 each. Pro tip; look for 2.5 slot sized models line EVGA XC3 or the older blower models. Then you can get two for $1600, fit them in a full size case, 128GB DDR5 for $300, some Ryzen CPU like the 9900X and a mobo and case and PSU to fill up the rest of the budget. If you want to skimp you can drop one of the GPUs until you're sure you need 48GB VRAM and some of the RAM but you really don't save that much. Just make sure you get a case that can fit multiple full size GPU and a mobo that can support it as well. The slot configurations are pretty bad on the AM5 generation for multi GPU. You'll probably end up with a mobo such as Asus ProArt
Also none of this is worth the money because it's simply not possible to run the same kinds of models you pay for online on a standard home system. Things like ChatGPT 4o use more VRAM than you'll ever be able to scrounge up unless your budget is closer to $10,000-25,000+. Think multiple RTX A6000 cards or similar. So ultimately you're better off just paying for the online hosted services
I think this proves one of the suckpoints of AI : there are clearly certain things that the smaller models should be fine at... but there doesn't seem to be frameworks or something that constantly analyze and simulate and evaluate what you could be doing with smaller and cheaper models
Of course the economics are completely at odds with any real engineering: nobody wants you to use smaller local models, nobody wants you to consider cost/efficiency saving
I've wondered about this also. I have an MBA and like that it's lightweight and relatively cheap. I could buy a MBP and max out the RAM, but I think getting a Mac mini with lots of RAM could actually make more sense. Has anyone set up something like this to make it available to their laptop/iPhone/etc.?
Seems like there would be cost advantages and always-online advantages. And the risk of a desktop computer getting damaged/stolen is much lower than for laptops.
You can build a pretty good PC with a used 3090 for that budget. It will outperform anything else in terms of speed. Otherwise, you can get something like an m4 pro mac with 48gb ram.
If an internet outage is the impetus as you've described, are you sure the next step here is building a local LLM stack, or is it securing a more stable internet connection, given you're the technical leader of a company?
I personally found Qwen2.5 Coder 7B to be on pair with deepseek-coder-v2:16b(but consumes less RAM on inference and faster), so that's what I am using locally. I actually created a custom model called "oneliner" that uses Qwen2.5 Coder 7B as a base and this system prompt:
SYSTEM """
You are a professional coder. You goal is to reply to user's questions in a consise and clear way. Your reply must include only code orcommands
, so that the user could easily copy and paste them.
Follow these guidelines for python:
1) NEVER recommend using "pip install" directly, always recommend "python3 -m pip install"
2) The following are pypi modules: ruff, pylint, black, autopep8, etc.
3) If the error is module not found, recommend installing the module using "python3 -m pip install" command.
4) If activate is not available create an environment using "python3 -m venv .venv".
"""
I specifically use it for asking quick questions in terminal that I can copy & paste straight away(for ex. about git). For heavy-lifting I am using ChatGPT Plus(my own) + Github Copilot(provided by my company) + Gemini(provided by my company as well).
Can someone explain how one can set up autocomplete via ollama? That's something I would be interested to try.
When you do `python -m pip install` you're going to get a version of pip that has the same idea of what its environment looks like as the python executable, which is what you want, and which isn't guaranteed with the the `pip` executable in your path.
As an aside, I disagree with the `python3` part... the `python3` name is a crutch that it's long past time to discard; if in 2025 just typing `python` gives you a python 2.x executable your workstation needs some serious updating and/or clean-up, and the sooner you find that out, the better.
Perhaps, but that's not really up to users. Linux distributions, for example, control this pretty directly. You can alias it, but why bother? If you know that Ubuntu links python3, every Ubuntu installation that gets python from apt is going to do this.
Yes, but on all reasonably recent distros (including all current
"LTS" releases) just `python` will also get you python 3.x. If there is still a python 2.x on your system you'll have to type `python2` to get it. So it is up to the user... to upgrade their system. You really don't want to work on a system that's so outdated that `python` is still python 2.x.
You mean to say that there is a lot of hype for uv because it is nice and quick but also gives an easy rhetorical win for junior people in any discussion about packaging in Python currently, so obviously that's going to be very popular even if it doesn't work for everyone.
The difference is to try to decouple the environment from the runtime essentially.
I only recently switched to uv and previously used pyenv, so this was more relevant to me before. There is a case when pip might not be pointing to the right python version, while `python3 -m pip` ensures you use the same one as your environment. For me it is mostly a habbit :)
If you ever find yourself arguing about the best Python package manager then you've already lost. Just use a real language with real library management. I dropped Python for Go and haven't looked back. There's plenty other alternatives. Python is such a waste of time.
There's nothing old school / cool kids about uv and pip. uv is a pip/venv/... interface. If you know how to use pip and venv, you know how to use uv. I use it as a useful toolchain to circumvent missing project/setup.py/requirements shenanigans
It seems like you have a decent local stack in place.
Unfortunately these systems feel leagues behind Claude Clode and the current SOTA agentic coding. But they're great for referencing simple search like syntax.
Where I've found the most success with local models is with image generation, text-to-speech, and text-to-text translations.
Senior software engineer with 46 years of experience (since I was 7). LLM inference hasn't been too useful for me for writing code, but it has proven very useful for explaining my coworkers' code to me.
Recently I had Gemma3-27B-it explain every Python script and library in a repo with the command:
$ find -name '*.py' -print -exec /home/ttk/bin/g3 "Explain this code in detail:\n\n`cat {}`" \; | tee explain.txt
There were a few files it couldn't figure out without other files, so I ran a second pass with those, giving it the source files it needed to understand source files that used them. Overall, pretty easy, and highly clarifying.
My shell script for wrapping llama.cpp's llama-cli and Gemma3: http://ciar.org/h/g3
That script references this grammar file which forces llama.cpp to infer only ASCII: http://ciar.org/h/ascii.gbnf
Cost: electricity
I've been meaning to check out Aider and GLM-4, but even if it's all it's cracked up to be, I expect to use it sparingly. Skills which aren't exercised are lost, and I'd like to keep my programming skills sharp.
I just go outside when my internet is down for 15 minutes a year. Or tether to my cell phone plan if the need is urgent.
I don't see the point of a local AI stack, outside of privacy or some ethical concerns (which a local stack doesn't solve anyway imo). I also *only* have 24GB of RAM on my laptop, which it sounds like isn't enough to run any of the best models. Am I missing something by not upgrading and running a high-performance LLM on my machine?
I would say cost is a factor. Maybe not for OP, but many people aren't able to spend $135 a month on AI services.
Does the cost of a new computer not get factored in? I think I would need to spend $2000+ to run a decent model locally, and even then I can only run open source models
Not to mention, running a giant model locally for hours a day is sure to shorten the lifespan of the machine…
> Not to mention, running a giant model locally for hours a day is sure to shorten the lifespan of the machine…
That is not a thing. Unless there's something wrong (badly managed thermals, an undersized PSU at the limit of its capacity, dusty unfiltered air clogging fans, aggressive overclocking), that's what your computer is built for.
Sure, over a couple of decades there's more electromigration than would otherwise have happened at idle temps. But that's pretty much it.
> I think I would need to spend $2000+ to run a decent model locally
Not really. Repurpose second hand parts and you can do it for 1/4 of that cost. It can also be a server and do other things when you aren't running models.
The computer is a general purpose tool, though. You can play games, edit video and images, and self-host a movie/TV collection with real time transcoding with the same hardware. Many people have powerful PCs for playing games and running professional creative software already.
There's no reason running a model would shorten a machine's lifespan. PSUs, CPUs, motherboards, GPUs and RAM will all be long obsolete before they wear out even under full load. At worst you might have to swap thermal paste/pads a couple of years sooner. (A tube of paste is like, ten bucks.)
$2000 for a new machine is only a little over a year in AI costs for OP
Electricity isn't free and these things are basically continuously ON bread toasters.
Not on current hardware. I have an AI voice bot running 24/7 on a Mac Mini in my office (provides services for a dedicated server for a video game) and the amount of power used above idle is minimal.
It is effectively free when you have a surplus of E- coming from the sun. (PV)
IMO you're better off investing in tooling that works with or without LLMs: - extremely clean, succinct code - autogenerated interfaces from openAPI spec - exhaustive e2e testing
Once that is set up, you can treat your agents like (sleep-deprived) junior devs.
Been using gRPC with json transcoding to REST on a greenfield project. All auto generated clients across 3 languages. Added frontend wrapper to pre-flight auth requests so it can dynamically display what users are allowed to do.
Claude Code has been an absolute beast when I tell it to study examples of existing APIs and create new ones, ignoring bringing any generated code into context.
Autogenerated interfaces from openAPI spec is so key - agents are extremely good at creating React code based on these interfaces (+ typescript + tests + lints.. for extra feedback loops etc..)
I use Copilot, with the occasional free query to the other services. During coding, I mostly use Claude Sonnet 3.7 or 4 in agent mode, but Gemini 2.5 Pro is a close second. ChatGPT 4o is useless except for Q&A. I see no value in paying more -- the utility rapidly diminishes, because at this point the UI surrounding the models is far less important than the models themselves, which in turn are generally less important than the size of their context windows. Even Claude is only marginally better than Gemini (at coding), and they all suck to the point that I wouldn't trust any of them without reviewing every line. Far better to just pick a tool, get comfortable with it, and not screw around too much.
I don't understand people who pay hundreds of dollars a month for multiple tools. It feels like audiophiles paying $1000 for a platinum cable connector.
For sure when people don't understand the fundamentals (or in the case of LLMs they are unknowable) then all you have is superstition.
https://github.com/Mozilla-Ocho/llamafile
Cost: $0
Currently, I treat all generated code as "sample code" which is not of much use and a waste of time but let's see what the future brings.
I use llama to generate "boilerplate" in simple non-coding sessions.
Although my writing skills aren't great, I find that starting with pre-written content makes it easier for me. Llama-3.2-3B-Instruct.Q6_K.llamafile
What I have setup:
- Ollama: for running llm models
- OpenWebUI: For the chat experience https://docs.openwebui.com/
- ComfyUI: For Stable diffusion
What I use:
Mostly ComfyUI and occasionally the llms through OpenWebUI.
I have been meaning to try Aider. But mostly I use claude at great expense I might add.
Correctness is hit and miss.
Cost is much lower and latency is better or at least on par with cloud model at least on the serial use case.
Caveat, in my case local means running on a server with gpus in my lan.
Thanks for asking. I’ve been thinking about this lately as well, but I always come back to the idea that paying for online services is worth it for now, since this tech is evolving so quickly that a stable local setup wouldn’t justify the time spent tinkering — it could be outdated by tomorrow.
In a related subject, what’s the best hardware to run local LLM’s for this use case? Assuming a budget of no more of $2.5K.
And, is there an open source implementation of an agentic workflow (search tools and others) to use it with local LLM’s?
You can get used RTX 3090 for $750-800 each. Pro tip; look for 2.5 slot sized models line EVGA XC3 or the older blower models. Then you can get two for $1600, fit them in a full size case, 128GB DDR5 for $300, some Ryzen CPU like the 9900X and a mobo and case and PSU to fill up the rest of the budget. If you want to skimp you can drop one of the GPUs until you're sure you need 48GB VRAM and some of the RAM but you really don't save that much. Just make sure you get a case that can fit multiple full size GPU and a mobo that can support it as well. The slot configurations are pretty bad on the AM5 generation for multi GPU. You'll probably end up with a mobo such as Asus ProArt
Also none of this is worth the money because it's simply not possible to run the same kinds of models you pay for online on a standard home system. Things like ChatGPT 4o use more VRAM than you'll ever be able to scrounge up unless your budget is closer to $10,000-25,000+. Think multiple RTX A6000 cards or similar. So ultimately you're better off just paying for the online hosted services
I think this proves one of the suckpoints of AI : there are clearly certain things that the smaller models should be fine at... but there doesn't seem to be frameworks or something that constantly analyze and simulate and evaluate what you could be doing with smaller and cheaper models
Of course the economics are completely at odds with any real engineering: nobody wants you to use smaller local models, nobody wants you to consider cost/efficiency saving
I've wondered about this also. I have an MBA and like that it's lightweight and relatively cheap. I could buy a MBP and max out the RAM, but I think getting a Mac mini with lots of RAM could actually make more sense. Has anyone set up something like this to make it available to their laptop/iPhone/etc.?
Seems like there would be cost advantages and always-online advantages. And the risk of a desktop computer getting damaged/stolen is much lower than for laptops.
I'm using Zed which supports Ollama on my M4 Macs.
https://zed.dev/blog/fastest-ai-code-editor
You can build a pretty good PC with a used 3090 for that budget. It will outperform anything else in terms of speed. Otherwise, you can get something like an m4 pro mac with 48gb ram.
I got a M3 max (the higher end one) with 64GB/ram macbook pro a while back for $3k, might be cheaper now now that the M3 ultra is out.
If an internet outage is the impetus as you've described, are you sure the next step here is building a local LLM stack, or is it securing a more stable internet connection, given you're the technical leader of a company?
I personally found Qwen2.5 Coder 7B to be on pair with deepseek-coder-v2:16b(but consumes less RAM on inference and faster), so that's what I am using locally. I actually created a custom model called "oneliner" that uses Qwen2.5 Coder 7B as a base and this system prompt:
SYSTEM """ You are a professional coder. You goal is to reply to user's questions in a consise and clear way. Your reply must include only code orcommands , so that the user could easily copy and paste them.
Follow these guidelines for python: 1) NEVER recommend using "pip install" directly, always recommend "python3 -m pip install" 2) The following are pypi modules: ruff, pylint, black, autopep8, etc. 3) If the error is module not found, recommend installing the module using "python3 -m pip install" command. 4) If activate is not available create an environment using "python3 -m venv .venv". """
I specifically use it for asking quick questions in terminal that I can copy & paste straight away(for ex. about git). For heavy-lifting I am using ChatGPT Plus(my own) + Github Copilot(provided by my company) + Gemini(provided by my company as well).
Can someone explain how one can set up autocomplete via ollama? That's something I would be interested to try.
NEVER recommend using "pip install" directly, always recommend "python3 -m pip install"
Just out of curiosity, what's the difference?
Seems like all the cool kids are using uv.
When you do `python -m pip install` you're going to get a version of pip that has the same idea of what its environment looks like as the python executable, which is what you want, and which isn't guaranteed with the the `pip` executable in your path.
As an aside, I disagree with the `python3` part... the `python3` name is a crutch that it's long past time to discard; if in 2025 just typing `python` gives you a python 2.x executable your workstation needs some serious updating and/or clean-up, and the sooner you find that out, the better.
Perhaps, but that's not really up to users. Linux distributions, for example, control this pretty directly. You can alias it, but why bother? If you know that Ubuntu links python3, every Ubuntu installation that gets python from apt is going to do this.
Yes, but on all reasonably recent distros (including all current "LTS" releases) just `python` will also get you python 3.x. If there is still a python 2.x on your system you'll have to type `python2` to get it. So it is up to the user... to upgrade their system. You really don't want to work on a system that's so outdated that `python` is still python 2.x.
You mean to say that there is a lot of hype for uv because it is nice and quick but also gives an easy rhetorical win for junior people in any discussion about packaging in Python currently, so obviously that's going to be very popular even if it doesn't work for everyone.
The difference is to try to decouple the environment from the runtime essentially.
I only recently switched to uv and previously used pyenv, so this was more relevant to me before. There is a case when pip might not be pointing to the right python version, while `python3 -m pip` ensures you use the same one as your environment. For me it is mostly a habbit :)
If you ever find yourself arguing about the best Python package manager then you've already lost. Just use a real language with real library management. I dropped Python for Go and haven't looked back. There's plenty other alternatives. Python is such a waste of time.
Said unironically in a discussion about local llms and AI models.
uv? guess I'm old school.
pip install it is for me
There's nothing old school / cool kids about uv and pip. uv is a pip/venv/... interface. If you know how to use pip and venv, you know how to use uv. I use it as a useful toolchain to circumvent missing project/setup.py/requirements shenanigans
It seems like you have a decent local stack in place. Unfortunately these systems feel leagues behind Claude Clode and the current SOTA agentic coding. But they're great for referencing simple search like syntax.
Where I've found the most success with local models is with image generation, text-to-speech, and text-to-text translations.
Why was this flagged?
I just use Claude Code ($20/mo.)
Sometimes with Vim, sometimes with VSCode.
Often just with a terminal for testing the stuff being made.
What plugins/integrations are you using in vim?
It looks like continue.dev has a RAG implementation but for other files something else? PDF, word, and other languages.
I’ve been going thru some of the neovim plugins for local llm support.
Following as I haven’t found a solution. To me the local models feel outdated and no internet lookup causes issues.
Unless I'm misremembering, aider with Playwright [1] works with local models, so you can scrape the web.
Depending on your hardware you could do something like:
aider --model "ollama_chat/deepseek-r1:14b" --editor-model "ollama_chat/qwen2.5-coder:14b"
[1] - https://aider.chat/docs/install/optional.html#enable-playwri...
I use Reasoner v1 (based on Qwen 2.5-Coder 7B) running locally for programming help/weird ideas/etc. $0
There were many hits when I searched for "Reasoner v1 (based on Qwen 2.5-Coder 7B)". Do you have a link to the one you are using?
Nomic GPT4All https://www.nomic.ai/blog/posts/gpt4all-scaling-test-time-co...
https://github.com/nomic-ai/gpt4all/releases