I hope they do well. AFAIK they’re training or finetuning an older LLaMA model, so performance might lag behind SOTA. But what really matters is that ETH and EPFL get hands-on experience training at scale. From what I’ve heard, the new AI cluster still has teething problems. A lot of people underestimate how tough it is to train models at this scale, especially on your own infra.
Disclaimer: I’m Swiss and studied at ETH. We’ve got the brainpower, but not much large-scale training experience yet. And IMHO, a lot of the “magic” in LLMs is infrastructure-driven.
No, the model has nothing do to with Llama. We are using our own architecture, and training from scratch. Llama also does not have open training data, and is non-compliant, in contrast to this model.
If you guys need help on GGUFs + Unsloth dynamic quants + finetuning support via Unsloth https://github.com/unslothai/unsloth on day 0 / 1, more than happy to help :)
So you're not going to use copyrighted data for training? That's going to be a disadvantage with respect to LLaMa and other well-known models, it's an open secret that everyone is using everything they can get their hands on.
Not sure about the Swiss laws, but the EU AI Act and the 2019/790 digital millennium directive it piggies back on the topic, does allow for training on copyrighted data as long as any opt-out mechanisms (e.g. robots.txt) are respected. AFAICT this LLM was trained by respecting those mechanisms (and as linked elsewhere they didn't find any practical difference in performance - note that there is an exception to allow ignoring the opt-out mechanisms for research purposes, so they could make that comparison).
Hey, really cool project, I’m excited to see the outcome.
Is there a blog / paper summarizing how you are doing it ?
Also which research group is currently working on it at eth ?
Can you comment on how the filtering impacted language coverage? E.g. finweb2 has 1800+ languages, but some with very little actual representation, while finweb2-hq has just 20 but each with a subdsantial data set.
(I'm personaly most interested in covering the 24 official EU languages)
we kept all 1800+ (script/language) pairs, not only the quality filtered ones. the question if a mix of quality filtered and not languages impacts the mixing is still an open question. preliminary research (Section 4.2.7 of https://arxiv.org/abs/2502.10361 ) indicates that quality filtering can mitigate the curse of multilinguality to some degree, so facilitate cross-lingual generalization, but it has to be seen how strong this effect is on larger scale
Imo, a lot of the magic is also dataset driven, specifically the SFT and other fine tuning / RLHF data they have. That's what has separated the models people actually use from the also-rans.
I agree with everything you say about getting the experience, the infrastructure is very important and is probably the most critical part of a sovereign LLM supply chain. I would hope there will also be enough focus on the data, early on, that the model will be useful.
When I read "from scratch", I assume they are doing pre-training, not just finetuning, do you have a different take? Do you mean it's normal Llama architecture they're using?
I'm curious about the benchmarks!
The infra does become pretty complex to get a SOTA LLM trained. People assume it's as simple as loading up the architecture and a dataset + using something like Ray. There's a lot that goes into designing the dataset, the eval pipelines, the training approach, maximizing the use of your hardware, dealing with cross-node latency, recovering from errors, etc.
But it's good to have more and more players in this space.
I wonder if the reason for these results is that any data on the internet is already copied to other locations by actors who ignore crawling opt-outs. So, even if they respect all web crawling opt-outs, they are still effectively copying the data because someone else did not respect it who does not include an opt-out.
Yes this is an interesting question. In our arxiv paper [1] we did study this for news articles, and also removed duplicates of articles (decontamination). We did not observe an impact on the downstream accuracy of the LLM, in the case of news data.
Is there not yet a Source where the web has already been scraped and souped down to just the text? It would seem someone would have created such a thing in order to save LLM training from having to reinvent the wheel.
I understand the web is a dynamic thing but still it would seem to be useful on some level.
No performance degradation on training metrics except for the end user. At the end of the day users and website owners have completely orthogonal interests. Users want answers and content, website owners want attention so they can upsell/push ads. You can only serve one master.
You don't. You bypass them with crawlers and don't reveal your training data. And this is exactly why open source models can't surpass open weight models.
> And this is exactly why open source models can't surpass open weight models.
It is a fair point, but how strong of a point it is remains to be seen, some architectures are better than others, even with the same training data, so not impossible we could at one point see some innovative architectures beating current proprietary ones. It would probably be short-lived though, as the proprietary ones would obviously improve in their next release after that.
Maybe the missing data makes it 3% worse but the architecture is 5% better. Or your respect for robots.txt gets you more funding and you gain a 4% advantage by training longer.
Don't focus too much on a single variable, especially when all the variables have diminishing returns.
It is logically impossible for a LLM to, for example, to know that fooExecute() takes two int arguments if the documentation is blocked by robots.txt and there are no examples of fooExecute() usage in the wild, don't you agree?
I agree, but also think it's less important. I don't want a big fat LLM that memorized every API out there, and as soon as the API changed, the weights have to updated. I like the current approach of Codex (and similar) where they can look up the APIs they need to use as they're doing the work instead, so same weights will continue to work no matter how much the APIs change.
Sure, the model would not “know” about your example, but that’s not the point; the penultimate[0] goal is for the model to figure out the method signature on its own just like a human dev might leverage her own knowledge and experience to infer that method signature. Intelligence isn’t just rote memorization.
I don't think a human dev can divine a method signature and effects in the general case either. Sure the add() function probably takes 2 numbers, but maybe it takes a list? Or a two-tuple? How would we or the LLM know without having the documentation? And yeah sure the LLM can look at the documentation while being used instead of it being part of the training dataset, but that's strictly inferior for practical uses, no?
I'm not sure if we're thinking of the same field of AI development. I think I'm talking about the super-autocomplete with integrated copy of all of digitalized human knowledge, while you're talking about trying to do (proto-)AGI. Is that it?
> Sure the add() function probably takes 2 numbers, but maybe it takes a list? Or a two-tuple? How would we or the LLM know without having the documentation?
You just listed possible options in the order of their relative probability. Human would attempt to use them in exactly that order
I think that the Allen Institute for Artificial Intelligence OLMo models are also completely open:
OLMo is fully open
Ai2 believes in the power of openness to build a future where AI is accessible to all. Open weights alone aren’t enough – true openness requires models to be trained in the open with fully open access to data, models, and code.
The open training data is a huge differentiator. Is this the first truly open dataset of this scale? Prior efforts like The Pile were valuable, but had limitations. Curious to see how reproducible the training is.
> The model will be fully open: source code and weights will be publicly available, and the training data will be transparent and reproducible
This leads me to believe that the training data won’t be made publicly available in full, but merely be “reproducible”. This might mean that they’ll provide references like a list of URLs of the pages they trained on, but not their contents.
Yeah, I suspect you're right. Still, even a list of URLs for a frontier model (assuming it does turn out to be of that level) would be welcome over the current situation.
This is an interesting problem that has various challenges - currently most tokenization solutions where trainees using hype pair encoding where the most commonly seen combinations of letters were being selected to be a mapping. This meant that the majority of tokenization was English mappings meaning your LLM had a better tokenization of English compared to other languages it was being trained on.
Sure, but usually you teach something that is inherently useful, or can be applied to some sort of useful endeavor. In this case I think it's fair to ask what the collision of two bubbles really achieves, or if it's just a useful teaching model, what it can be applied to.
The model will be released in two sizes — 8 billion and 70 billion parameters [...]. The 70B version will rank among the most powerful fully open models worldwide. [...] In late summer, the LLM will be released under the Apache 2.0 License.
This seems like the equivalent of a university designing an ICE car...
What does anyone get out of this when we have open weight models already ?
Are they going to do very innovative AI research that companies wouldn't dare try/fund? Seems unlikely ..
Is it a moonshot huge project that no single company could fund..? Not that either
If it's just a little fun to train the next generation of LLM researchers.. Then you might as well just make a small scale toy instead of using up a super computer center
Including how it was trained, what data was used, how training data was synthesized, how other models were used etc. All the stuff that is kept secret in case of llama, deepseek etc.
Pretty proud to see this at the top of HN as a Swiss (and I know many are lurking here!). These two universities produce world-class founders, researchers, and engineers. Yet, we always stay in the shadow of the US. With our top-tier public infrastructure, education, and political stability (+ neutrality), we have a unqiue opportunity to build something exceptional in the open LLM space.
“ Open LLMs are increasingly viewed as credible alternatives to commercial systems, most of which are developed behind closed doors in the United States or China”
It is obvious that the companies producing big LLMs today have the incentive to try to enshitify them. Trying to get subscriptions at the same time as trying to do product placement ads etc. Worse, some already have political biases they promote.
It would be wonderful if a partnership between academia and government in Europe can do a public good search and AI that endeavours to serve the user over the company.
Yes but it’s a very complicated service to deliver. Even if they train great models, they likely will not operationalize them for inference. Those will still be private actors, and the incentives to enshittify will be the same. Also, for AI generally the incentives is much higher than last tech generation, due to cost of running these things. Basically, the free services where you’re the product must aggressively extract value out of you in order to make a profit.
Will be interested to see how this model responds to currently unresolvable issues in physics. Is it an open or a closed world mentality and/or a conditioned disclaimer which encourages progress?
Sometimes ago there was a Tom Scott video about the fasted accelerating car in the world, developed by a team with a vast majority of student. One remark stayed with me: "the goal is not to build a car, but to build engineer".
In that regard it's absolutely not a waste of public infra just like this car was not a waste.
The announcement was at the International Open-Source LLM Builders Summit held this week in Switzerland. Is it so strange that they announced what they are doing and the timeline?
Funding? Deeply biasing European uses to publicly-developed European LLMs (or at least not American or Chinese ones) would make a lot of sense. (Potentially too much sense for Brussels.)
I hope they do well. AFAIK they’re training or finetuning an older LLaMA model, so performance might lag behind SOTA. But what really matters is that ETH and EPFL get hands-on experience training at scale. From what I’ve heard, the new AI cluster still has teething problems. A lot of people underestimate how tough it is to train models at this scale, especially on your own infra.
Disclaimer: I’m Swiss and studied at ETH. We’ve got the brainpower, but not much large-scale training experience yet. And IMHO, a lot of the “magic” in LLMs is infrastructure-driven.
No, the model has nothing do to with Llama. We are using our own architecture, and training from scratch. Llama also does not have open training data, and is non-compliant, in contrast to this model.
Source: I'm part of the training team
If you guys need help on GGUFs + Unsloth dynamic quants + finetuning support via Unsloth https://github.com/unslothai/unsloth on day 0 / 1, more than happy to help :)
absolutely! i've sent you a linkedin message last week. but here seems to work much better, thanks a lot!
So you're not going to use copyrighted data for training? That's going to be a disadvantage with respect to LLaMa and other well-known models, it's an open secret that everyone is using everything they can get their hands on.
Good luck though, very needed project!
Not sure about the Swiss laws, but the EU AI Act and the 2019/790 digital millennium directive it piggies back on the topic, does allow for training on copyrighted data as long as any opt-out mechanisms (e.g. robots.txt) are respected. AFAICT this LLM was trained by respecting those mechanisms (and as linked elsewhere they didn't find any practical difference in performance - note that there is an exception to allow ignoring the opt-out mechanisms for research purposes, so they could make that comparison).
Hey, really cool project, I’m excited to see the outcome. Is there a blog / paper summarizing how you are doing it ? Also which research group is currently working on it at eth ?
Thanks for clarifying! I wish you all the best luck!
L3 has open pretraining data, it's just not official for obvious legal reasons: https://huggingface.co/datasets/HuggingFaceFW/fineweb
Are you using dbpedia?
no. the main source is fineweb2, but with additional filtering for compliance, toxicity removal, and quality filters such as fineweb2-hq
Thx for engaging here.
Can you comment on how the filtering impacted language coverage? E.g. finweb2 has 1800+ languages, but some with very little actual representation, while finweb2-hq has just 20 but each with a subdsantial data set.
(I'm personaly most interested in covering the 24 official EU languages)
we kept all 1800+ (script/language) pairs, not only the quality filtered ones. the question if a mix of quality filtered and not languages impacts the mixing is still an open question. preliminary research (Section 4.2.7 of https://arxiv.org/abs/2502.10361 ) indicates that quality filtering can mitigate the curse of multilinguality to some degree, so facilitate cross-lingual generalization, but it has to be seen how strong this effect is on larger scale
Imo, a lot of the magic is also dataset driven, specifically the SFT and other fine tuning / RLHF data they have. That's what has separated the models people actually use from the also-rans.
I agree with everything you say about getting the experience, the infrastructure is very important and is probably the most critical part of a sovereign LLM supply chain. I would hope there will also be enough focus on the data, early on, that the model will be useful.
When I read "from scratch", I assume they are doing pre-training, not just finetuning, do you have a different take? Do you mean it's normal Llama architecture they're using? I'm curious about the benchmarks!
The infra does become pretty complex to get a SOTA LLM trained. People assume it's as simple as loading up the architecture and a dataset + using something like Ray. There's a lot that goes into designing the dataset, the eval pipelines, the training approach, maximizing the use of your hardware, dealing with cross-node latency, recovering from errors, etc.
But it's good to have more and more players in this space.
"respecting web crawling opt-outs during data acquisition produces virtually no performance degradation"
Great to read that!
I wonder if the reason for these results is that any data on the internet is already copied to other locations by actors who ignore crawling opt-outs. So, even if they respect all web crawling opt-outs, they are still effectively copying the data because someone else did not respect it who does not include an opt-out.
Yes this is an interesting question. In our arxiv paper [1] we did study this for news articles, and also removed duplicates of articles (decontamination). We did not observe an impact on the downstream accuracy of the LLM, in the case of news data.
[1] https://arxiv.org/abs/2504.06219
My guess is that it doesn't remove that much of the data, and the post-training data (not just randomly scraped from the web) probably matters more
Is there not yet a Source where the web has already been scraped and souped down to just the text? It would seem someone would have created such a thing in order to save LLM training from having to reinvent the wheel.
I understand the web is a dynamic thing but still it would seem to be useful on some level.
Common Crawl, maybe?
No performance degradation on training metrics except for the end user. At the end of the day users and website owners have completely orthogonal interests. Users want answers and content, website owners want attention so they can upsell/push ads. You can only serve one master.
> Users want answers and content, website owners want attention so they can upsell/push ads. You can only serve one master
How are you going to serve users if web site owners decide to wall their content? You can't ignore one side of the market.
You don't. You bypass them with crawlers and don't reveal your training data. And this is exactly why open source models can't surpass open weight models.
> And this is exactly why open source models can't surpass open weight models.
It is a fair point, but how strong of a point it is remains to be seen, some architectures are better than others, even with the same training data, so not impossible we could at one point see some innovative architectures beating current proprietary ones. It would probably be short-lived though, as the proprietary ones would obviously improve in their next release after that.
How can open source models respectful of robots.txt possibly perform equally if they are missing information that the other models have access to?
Maybe the missing data makes it 3% worse but the architecture is 5% better. Or your respect for robots.txt gets you more funding and you gain a 4% advantage by training longer.
Don't focus too much on a single variable, especially when all the variables have diminishing returns.
How can we possibly find out without trying?
It is logically impossible for a LLM to, for example, to know that fooExecute() takes two int arguments if the documentation is blocked by robots.txt and there are no examples of fooExecute() usage in the wild, don't you agree?
I agree, but also think it's less important. I don't want a big fat LLM that memorized every API out there, and as soon as the API changed, the weights have to updated. I like the current approach of Codex (and similar) where they can look up the APIs they need to use as they're doing the work instead, so same weights will continue to work no matter how much the APIs change.
Sure, the model would not “know” about your example, but that’s not the point; the penultimate[0] goal is for the model to figure out the method signature on its own just like a human dev might leverage her own knowledge and experience to infer that method signature. Intelligence isn’t just rote memorization.
[0] the ultimate, of course, being profit.
I don't think a human dev can divine a method signature and effects in the general case either. Sure the add() function probably takes 2 numbers, but maybe it takes a list? Or a two-tuple? How would we or the LLM know without having the documentation? And yeah sure the LLM can look at the documentation while being used instead of it being part of the training dataset, but that's strictly inferior for practical uses, no?
I'm not sure if we're thinking of the same field of AI development. I think I'm talking about the super-autocomplete with integrated copy of all of digitalized human knowledge, while you're talking about trying to do (proto-)AGI. Is that it?
> Sure the add() function probably takes 2 numbers, but maybe it takes a list? Or a two-tuple? How would we or the LLM know without having the documentation?
You just listed possible options in the order of their relative probability. Human would attempt to use them in exactly that order
this is what this paper tries to answer: https://arxiv.org/abs/2504.06219 the quality gap is surprisingly small between compliant and not
ETH Zurich is doing so many amazing things that I want to go study there. Unbelievable how many great people are coming from that university
Is this setting the bar for dataset transparency? It seems like a significant step forward. Assuming it works out, that is.
They missed an opportunity though. They should have called their machine the AIps (AI Petaflops Supercomputer).
I think that the Allen Institute for Artificial Intelligence OLMo models are also completely open:
OLMo is fully open
Ai2 believes in the power of openness to build a future where AI is accessible to all. Open weights alone aren’t enough – true openness requires models to be trained in the open with fully open access to data, models, and code.
https://allenai.org/olmo
I am a simple man, I see AI2, I upvote.
Smollm is also completely open as far as I know
The open training data is a huge differentiator. Is this the first truly open dataset of this scale? Prior efforts like The Pile were valuable, but had limitations. Curious to see how reproducible the training is.
> The model will be fully open: source code and weights will be publicly available, and the training data will be transparent and reproducible
This leads me to believe that the training data won’t be made publicly available in full, but merely be “reproducible”. This might mean that they’ll provide references like a list of URLs of the pages they trained on, but not their contents.
Well, when the actual content is 100s of terabytes big, providing URLs may be more practical for them and for others.
The difference between content they are allowed to train on vs. being allowed to distribute copies of is likely at least as relevant.
That wouldn't seem reproducible if the content at those URLs changes. (Er, unless it was all web.archive.org URLs or something.)
This is a problem with the Web. It should be easier to download content like it was updating a git Repo.
Yeah, I suspect you're right. Still, even a list of URLs for a frontier model (assuming it does turn out to be of that level) would be welcome over the current situation.
Yup, it’s not a dataset packaged like you hope for here, as it still contains traditionally copyrighted material
Yeah, that's what "democratizing AI" means.
I wonder if multilingual llms are better or worse compared a single language model
This is an interesting problem that has various challenges - currently most tokenization solutions where trainees using hype pair encoding where the most commonly seen combinations of letters were being selected to be a mapping. This meant that the majority of tokenization was English mappings meaning your LLM had a better tokenization of English compared to other languages it was being trained on.
C.f. https://medium.com/@biswanai92/understanding-token-fertility...
The press release talks a lot about how it was done, but very little about how capabilities compare to other open models.
It's a university, teaching the 'how it's done' is kind of the point
Sure, but usually you teach something that is inherently useful, or can be applied to some sort of useful endeavor. In this case I think it's fair to ask what the collision of two bubbles really achieves, or if it's just a useful teaching model, what it can be applied to.
The model will be released in two sizes — 8 billion and 70 billion parameters [...]. The 70B version will rank among the most powerful fully open models worldwide. [...] In late summer, the LLM will be released under the Apache 2.0 License.
We'll find out in September if it's true?
Yeah, I was thinking more of a table with benchmark results
I hope DeepSeek R2, but I fear Llama 4.
This seems like the equivalent of a university designing an ICE car...
What does anyone get out of this when we have open weight models already ?
Are they going to do very innovative AI research that companies wouldn't dare try/fund? Seems unlikely ..
Is it a moonshot huge project that no single company could fund..? Not that either
If it's just a little fun to train the next generation of LLM researchers.. Then you might as well just make a small scale toy instead of using up a super computer center
Why do you think it's about money? IMO it's about much more than that, like independence and actual data freedom trough reproductive LLMs
That it will actually be open and reproducible?
Including how it was trained, what data was used, how training data was synthesized, how other models were used etc. All the stuff that is kept secret in case of llama, deepseek etc.
This model will be one of the few open models where the training data is also open which makes it ideal for fine tuning.
Pretty proud to see this at the top of HN as a Swiss (and I know many are lurking here!). These two universities produce world-class founders, researchers, and engineers. Yet, we always stay in the shadow of the US. With our top-tier public infrastructure, education, and political stability (+ neutrality), we have a unqiue opportunity to build something exceptional in the open LLM space.
I work with EPFL alumni. Brilliant minds.
The article says
“ Open LLMs are increasingly viewed as credible alternatives to commercial systems, most of which are developed behind closed doors in the United States or China”
It is obvious that the companies producing big LLMs today have the incentive to try to enshitify them. Trying to get subscriptions at the same time as trying to do product placement ads etc. Worse, some already have political biases they promote.
It would be wonderful if a partnership between academia and government in Europe can do a public good search and AI that endeavours to serve the user over the company.
Yes but it’s a very complicated service to deliver. Even if they train great models, they likely will not operationalize them for inference. Those will still be private actors, and the incentives to enshittify will be the same. Also, for AI generally the incentives is much higher than last tech generation, due to cost of running these things. Basically, the free services where you’re the product must aggressively extract value out of you in order to make a profit.
Looking forward to proof test it.
nice
Use case for science and code LLMs: Superhydrodynamic gravity (SQR / SQG, )
LLMs do seem to favor general relativity but probably would've favored classical mechanics at the time given the training corpora.
Not-yet unified: Quantum gravity, QFT, "A unified model must: " https://news.ycombinator.com/item?id=44289148
Will be interested to see how this model responds to currently unresolvable issues in physics. Is it an open or a closed world mentality and/or a conditioned disclaimer which encourages progress?
What are the current benchmarks?
From https://news.ycombinator.com/item?id=42899805 re: "Large Language Models for Mathematicians" (2023) :
> Benchmarks for math and physics LLMs: FrontierMath, TheoremQA, Multi SWE-bench: https://news.ycombinator.com/item?id=42097683
Multi-SWE-bench: A Multi-Lingual and Multi-Modal GitHub Issue Resolving Benchmark: https://multi-swe-bench.github.io/
Add'l LLM benchmarks and awesome lists: https://news.ycombinator.com/item?id=44485226
Microsoft has a new datacenter that you don't have to keep adding water to; which spares the aquifers.
How to use this LLM to solve energy and sustainability problems all LLMs exacerbate? Solutions for the Global Goals, hopefully
gross use of public infrastructure
Sometimes ago there was a Tom Scott video about the fasted accelerating car in the world, developed by a team with a vast majority of student. One remark stayed with me: "the goal is not to build a car, but to build engineer".
In that regard it's absolutely not a waste of public infra just like this car was not a waste.
It even used green power. Literally zero complains or outcry from the public yet. Guess we like progress, especially if it helps independence.
I literally cant fault this, even steelmanning anti AI positions. What makes you say that?
Why would you announce this without a release? Be honest.
The announcement was at the International Open-Source LLM Builders Summit held this week in Switzerland. Is it so strange that they announced what they are doing and the timeline?
The cliché (at least on my side of the Alps) is that people in Switzerland like to take theiiiir tiiiime.
"Move as quickly as possible, but as slowly as necessary."
Funding? Deeply biasing European uses to publicly-developed European LLMs (or at least not American or Chinese ones) would make a lot of sense. (Potentially too much sense for Brussels.)