Other fully open LLMs include Allen AI's OLMo 3.1 and MBZUAI's K2 Think V2, both of which have released their full training pipelines and datasets.
Nvidia Nemotron is also an open training source model, though a portion of its dataset remains proprietary.
Quoting lambda's comment:
> Note that the Nemotron models are generally stronger than Olmo and K2 Think V2 (according to Artificial Analysis benchmarks), and there is a lot of overlap in their datasets (lots of datasets are based on the same sources with different filtering, Olmo and K2 Think V2 both have used some Nemotron datasets).
> But yeah, Nemotron is a modern and fairly capable LLM, even the 122b is more capable than Deepseek R1 (a 671b model) on most benchmarks, and there's also the recently released 550b Ultra now.
Allen AI do not get enough love. They are doing GenAI how it should have always been done.
In fact, if the frontier companies had taken their approach, it would have started much slower, but I think we would be far more advanced by 2035. Instead we have a majority of society that wants to see AI fail.
I like the idea, and it has become more pressing that everyone outside the US think about tech sovereignty because the US has become an unsafe place to keep your data, but the impression I get from Apertus is that it moves at the speed of a committee. I have no expectation they'll deliver a competitive model. At least, not competitive with current models. Maybe competitive with models a year ago (though they haven't even done that yet, right?).
"the US has become an unsafe place to keep your data"
I empathize with this but curious what would make any other country a better safehaven for your data? I personally like the EU's approach to data safeguards, but are there other locales/data protections you have in mind that would keep your data "safe".
From a legal perspective the US may be safer than other places if the US is the one seeking your data. The US doesn't need legal process to authorize digging into your foreign server.
From a practical perspective, I'm not sure any servers are safe anywhere.
Iceland and Switzerland are probably the best places to keep your data safe. I'd put Norway, Sweden, Germany, and the Netherlands after that, though I don't have much specifics on how good they are at privacy these days.
Illegal tariffs, executive usurping congress power of the purse, Noem funding herself and others with a commercial from an unknown entity with tax payer money, people in ICE/FBI handing over undisclosed unaccounted money in brown bags, insider trading is rampant, using funds in appropriately to fly girlfriend places that isn't official business, illegally using private money to fund public projects, taking bribes from foreign nations like jets and such violating emulation clauses, passing no bid contracts to people you know, using the pardon power inappropriately to pardon crypto scammers and other white collar crimes, moving notorious Epstein related criminals to a low security prison without going through the courts, avoiding justice for sex crimes, using the DOJ as a political cudgel, and the list goes on.
It is a commonly accepted "fact" right now, outside the US, that the US is not to be trusted (right now), due to some orange guy, and his mates, manipulating markets, running their mouths, doing all kinds of criminal and/or infantile shit.
I'd say there is quite a bit of evidence for this all around.
For a model that claims to focus on many languages, it's quite unreliable when it comes to simple questions like "how to say X in language Y" or "how to conjugate verb X in language Y". It keeps hallucinating words that do not exist, and when corrected, it only hallucinates a new lie.
By far the most impactful product of the Apretus project are the people. To quote a memorable line from Dominique Paul (https://www.thisiscrispin.com/):
> What most people miss IMO is that this is not a team who is doing this for the fourth time like virtually any other LLM provider and who could learn from its own past experiences. I bet if the team would do another model training they could get way better results at one fourth of the costs.
Sovereign AI is not about using just one model. It's about using the right model for the right job, and getting them to talk through the solution TOGETHER before presenting the answer.
I'm curious to know what stuff like this means for cohere? Their whole value prop is Sovereign AI. It seems they spent a lot of money developing models but own none of their own infra, what is the point of a country spending a lot of money on coheres solutions when stuff like this is becoming increasingly available and usable? Feels like I must be missing something here??
The previous version of this model has been pretty bad, but claimed to adhere to copyright laws. However, based on my testing, that's not true either. So in my view this is completely useless.
As long as the following remains true, this release ends up a bigger contribution to science at large than most other models trained "behind closed doors":
> Fully open model: open weights + open data + full training details including all data and training recipes
It's good that there is a movement for open LLMs, but it's not where the battleground is right now. The battleground is local vs service LLMs, and we are losing that battle badly despite all the software being here now and viable, entirely because UX sucks.
How many normal people do you know who use "ChatGPT"? A lot, probably.
How many even know what "Gemma" is, let alone have downloaded llama.cpp, a GGUF file from Hugginface, and run "llama-server" from a text console with all the correct command arguments? How many are thinking about this use case when speccing out their next computer? Where is the breathless marketing copy boasting x tok/s?
"Normal people" have never bothered to host their own: photos, music, videos, documents, comunications, etc. To the point that for many their computer is essentially a thin client into someone else's server. Why would we think this same people would care about "personal" inference?
Normal people can go open an account at DeepSeek or Xiaomi and chat away for free. Or, for that matter, a couple other models like z.ai's (GLM-5.2 isn't in the free tier, though, but neither is GPT-5.5-Pro), or Qwen, which does have 3.7-Max for free with no account on their chatbot interface.
Yes, I realise this isn't "running a local model", but it's using models that can be grabbed and run locally. For my pipelines, I feel far more confidence when I use an open model (even one like GLM-5.2 that would be expensive for me to run) since I have a backup plan if the hosted/cloud option becomes unworkable for me. If that happens to me with Opus, I have zero options.
Google Edge Gallery is turn key for people and on the device most people chatgpt on. Just like with most Google Stuff “edge gallery” is maybe the worst name possible for “run AI on your phone”!
They may not right now, but the whole point of Microsoft's Copilot+ PC standard (even though it's somewhat anemic) is to run models locally. Apple Silicon with enough unified memory is capable. Not to mention modern iPhones and Pixels have fairly capable NPUs and routinely run local models. So, we may not be to the point where most normal people have the hardware to run local models, but it is rapidly approaching.
They have it, we just haven’t enabled them. The smart model with a chat box is the wrong abstraction for local. Ideally we would have it built into applications as a clear and easy to use opt-in feature. Like allowing a user to index a folder on their hard drive and then search it semantically via embeddings. You could do that on fairly low end hardware these days. Like 2GB of RAM with any processor made within the last 10 years.
it's funny because i made this thing (called enough) that aims to make it easy for non-technical people to get up and running with local models quickly, but it is impossible to figure out how to break through the noise. every thread and comment like this breaks my heart a lil bit
Better UX does not buy you a datacenter farm to train state of the art cutting edge models. Right now the only people who can do that are the technobility class.
It does not, but it might encourage more people to care. Worrying about training is a luxury when you are starting from a baseline of "OpenAI spies upon me and controls my access". Let's focus on getting every Tom, Dick and Harry 1) on board with LLMs, because they're happening, 2) habitually using local software.
> we are losing that battle badly despite all the software being here now and viable, entirely because UX sucks.
Yep. I'm an old time Linux sysadmin, but I am COMPLETELY baffled as to what I can or cannot run on my 32GB R9700 with 128GB main CPU memory.
If I want something Claude or Codex like what do I use that would be useful? If I want a chat system, what do I use? Images--apparently ComfyUI for setup but after that what do I do?
I don't even mind spinning up something in the cloud for a bit, but I need to know how I'm going to get data up and down without racking up massive bandwidth charges.
I'd love to do some tinkering, but the field is moving so fast and so full of charlatans that cleaning the dross out is almost impossible.
I think a problem with open-weight models is that while you can improve them, you are not going to create the next generation of LLMs by fine-tuning. We are at the mercy of frontier labs for access to SOTA LLMs. For example, Anthropic recently started requiring identity verification for Claude [0], same for OpenAI [1].
If one day China's distillation labs stop releasing their LLMs as open-weight, I doubt American labs will continue to release free LLM weights without that competition.
That's where fully open pipelines shine: they enable the community to create the next generation of SOTA LLMs. That is the only way LLMs truly become sovereign.
This notion that Chinese labs are merely distilling frontier models is quite an unwarranted slur. Those labs have published WAY more useful research than US labs on RL techniques, novel model architectures, training pipelines, etc. They have also hit intelligence-per-parameter densities that US labs have yet to attain.
Apart from that, merely training a model on outputs from another model, off policy and without the logits, doesn’t really work that well.
The Chinese labs know how to build frontier level models. GLM-5.2 shows that they no longer even need Nvidia chips to do it.
It's one of those lies people tell themselves to make themselves feel better. "Oh, they're just copying my stuff."
Chinese labs are basically just telling everyone, out in the open, what they're doing and how to do it, and the answer from American frontier labs is "Well, they couldn't possibly be getting the results they're getting without just distilling our models," and the American labs aren't even trying to do some of the stuff like DS's aggressive caching to get costs down.
But have they? I understand that the Chinese side is illuminated and the American side is dark. I disagree that the Chinese labs have created anything that isn't in an American research lab or production dc. Sure the Chinese have published their findings and not for nothing. But are they novel? Unlikely imo
They are doing ta tremendous amount of novel research where American AI companies have "war rooms" to study their papers and models and American labs publish next to nothing. They have to often do more with less. As an AI researcher, Chinese labs are doing tremendous benefit to science whereas some American companies (and I'm American) seem to think only they are able to do AI research responsibility (I've been working on neural networks for 25+ years). I'm pretty sure Fable sabotaged my research codebase (see the news stories about this).
I recently watched a video for one of these “Chinese Models” it kept insisting it was Claude when the user asked. Sorry, there’s no “slur” here but legit suspicion.
These anecdotes where someone gets the model to claim it is X model are meaningless. (Claude also has been known to claim it is Deepseek when asked in Chinese.)
As anyone who's tried to write an AGENTS.md that says "Place an Assisted-by: git trailer that contains the harness you're using:whatever model this is"; such a naive approach often results in a seemingly random model.
> We are at the mercy of frontier labs for access to SOTA LLMs
I disagree with this use of SOTA, and this topic is why.
Anthropic and OpenAI have “cutting-edge” models. These are beyond the state of the art but they are closed, secretive, hard to quantify.
The “state of the art” is open source, open weights models that can be inspected, studied, shared and critiqued, because that is what is meant by “the art” —- it is the knowledge and principles and evidence and materials available to all. The “state of the art” is the highest point of that.
I wish we could make this distinction and stop blessing two secretive, unverifiable loss-making companies with so much power.
(Putting that aside, I suspect — without evidence, mind you - that the endless march to solving models by making them bigger is not the solution anyway.)
SOTA LLMs is less important than cheap token and Chinese AI labs is releasing model that is only about 6-8 months behind American AI labs.
Chinese's model like GLM is getting better for coding task and its cheaper. Microsoft Github copilot have to switch billing to token based. the cost of AI have increased since agent come into play. whoever can offer cheaper token to do task will win.
even Microsoft is looking into Deepseek for cheap token.
Sorry but I think you’re requirement that something only be “the art” if any arbitrary person can critique it is off.
The frontier labs are working on the state of the art but it’s just art that you aren’t allowed to see.
Unfortunately.
It is work using the principles of the art, obviously.
But "state of the art" implies the highest state of general availability, not just in terms of access to some product, but of use of the ideas, concepts, methodologies etc.
Anthropic and OpenAI have "cutting edge" models; the state of the art is behind the cutting edge.
The state of the art is the best open source, open weights model available. More or less by definition.
I appreciate this distinction. The are multiple senses of SOTA and one that has been taking on greater mindshare is as a synonym of “the best available”. By rebasing on SOTA as generally available and understood versus cutting edge, which has limited distribution and leads the way, we expand the vocabulary we have available to describe what’s going on. Thanks.
That's an interesting and possibly useful distinction , but it seems unique to you. Spreading it as "We should categorize the AIs this way" would be a good argument.
But the way SOTA is generally understood by other users of the language, it refers to exactly the team, technology, & techniques defining the cutting edge in any field, regardless of the whether the technology & techniques are available outside of that team...
I use it extensively. It is not ready for agentic use, but as a generic driving model for RAG use cases, it is pretty competent. You can build useful software with it.
As an opesource AI researcher with a lot of models and datasets on huggingface I am very appreciative of these types of project but we are ignoring the elephant in the room here ( or lack of )
How is this a real problem? Genuine question, because i don’t really understand the urgency of everyone buying up ram and gpus as prices for those skyrocket.
I can run the 8B version of this swiss-ai model on a ten year old GPU. For the larger one, $2000 consumer hardware can run it fine. Beyond that, there are plenty of places where time on a GPU can be rented, and if the model is good, there will be hardware to run it.
I'm mildly surprised that more people aren't using Nemo models for this reason. We've moved most of our processing to a combination of Nemo Ultra and Super, with some support for multi-model-specific tasks on Omni. The setup is working REALLY well for us, and I'm comfortable with the more measured pace of improvements. We work with many long-context problems, and the ecosystem is great.
There were a number of use cases where we needed to use Gemini (audio modality), and Ultra has been a VERY cost-effective alternative once we got through the nuances.
Nvidia Nemotron is also an open training source model, though a portion of its dataset remains proprietary.
Quoting lambda's comment:
> Note that the Nemotron models are generally stronger than Olmo and K2 Think V2 (according to Artificial Analysis benchmarks), and there is a lot of overlap in their datasets (lots of datasets are based on the same sources with different filtering, Olmo and K2 Think V2 both have used some Nemotron datasets).
> But yeah, Nemotron is a modern and fairly capable LLM, even the 122b is more capable than Deepseek R1 (a 671b model) on most benchmarks, and there's also the recently released 550b Ultra now.
https://news.ycombinator.com/item?id=48492439