Fully open models really need to be a big part of the AI future. That includes all source code, open training data, how it's organized, fed to the model, processed, etc. Until that becomes a thing you're always going to be left wondering what exactly lies underneath the closed model you are using, leaving open the possibility for societal manipulation.
Or, we just need to get this over with and declare any digital data findable via the internet to just be public property of everyone. Everything becomes public, besides stuff you keep locally, and there is no difference anymore, it's all just data anyone can use for whatever. A 1 year grace period for everyone to pull stuff off they don't want to be a part of this bright new open era, then we just scrap everything related to intellectual property, copyright and similar stupid stuff, and slap UBI on top of all of it for good measure.
I'd prefer to stay within the [hacker ethics](https://www.ccc.de/en/hackerethics), and protect private data. For non-private/personal data, sure. But individual people need their privacy protected.
You could sidestep it by running non-permissibly licensed training data that you purchased through an LLM. Legal attitude so far seems to be that this is transformative as long as it's not 1:1. The question on whether or not the end result is copyrightable of course remains controversial and inconsistent, but that question is also fairly irrelevent. You don't get more libre than public domain.
That's a fair amount of computational and labor overhead mind you, as you'll need to verify and prune the quality of your mountain of synthetic data, but certainly possible.
Though this assumes the legal system is a rational actor playing by the set of rules it claims to. In fact, I highly suspect you could get very unlucky and get an unfavorable ruling against you, because you stepped on a big pile of money's toes in the process of doing this.
Decentralized unstoppable storage, combined with decentralized unstoppable training, sorta like SETI for AI training. The seed of this tech already exists with IPFS and others like it.
We know (some? all?) of the big labs have skirted copyright laws at one point or another. Truly open models would just build on what is publicly available.
If the LLM/AI ecosystem starts actually needing some Person-To-Person (or maybe Agent-To-Agent?) payment system because things actually get smart enough to be useful autonomously, they're gonna need some way to send money/currency around. Depending on how banks will react to this need, we might see another return of digital currencies from the current winter.
Eventually we'll just construct 100% synthetic training data that can reliably reproduce pretrains and fine tunes.
The first broadly useful fully open source models will do this.
We already have open data / open code / open weights for some domain-specific cases, such as audio models trained on large open datasets, eg. Tacotron / LJSpeech from waaay back in the day, though that is certainly not SOTA anymore.
Distillation could possibly be considered an early case of this as raw AI outputs are themselves not copyrightable unless humans enrich, filter, or transform them. Granted, that does not handle the cases where the outputs are sufficiently similar to copyrighted original works.
But how much of that synthetic data still ultimately derives from non-open sources? You'd still have to ask what a clean room implementation ultimately is, depending on how granular or aggressive a large publisher wanted to get about it.
That said, I don't necessarily disagree with you. Talkie[1] presents an interesting case for it being at least possible to do this entirely on public domain material.
But even that used Claude somewhere in the course of its training pipeline (it's listed as a contributor on their GitHub), so again, how granular you want to get with that is still a question.
Why? Sure, I’d prefer it, too, but this is just another GNU/Linux vs. macOS situation: most of us would prefer the first, but actually get shit done on the latter.
we get shit done on the cloud with the former rather than the later
I personally find the analogy unconvincing, the UX dimension is completely different as I can use the same harness with any model; and the year of the linux desktop is coming soon (tm)
Open models can be used/changed for social manipulation too, by anyone, which scares a bunch of people, as opposed to the dark pattern manipulation from Big Ai/Tech
My quick review of the 3.7B model (because I was interested) is that it's not to be trusted for coding.
It failed my basic test I like to ask models and generated incorrect code. When prompted about the bug, it preceded to start hallucinating non-existent APIs. After doing that it got caught in a loop trying to desk check the solution that didn't work.
I'm glad to see some development in the space of "truly open" models that share training data and other recipes. As the costs for hardware fall over time (hopefully), we should see more possibility in fine-tuning and developing software to inspect the source training material.
Some other open models I'm aware of:
- OLMo
- Apertus
- Soofi
- OpenEuroLLM
- llm-jp
OLMo is perhaps the most famous, and their Dolma training corpus has been reused in other projects. It looks like the K2 training materials haven't been released yet, but I'm interested to see what they did for training "long-horizon agentic tasks". I'm aware of SWE-smith + SWE-gym but I'm guessing there's a lot more out there now.
I'm no expert, which is part of why these projects excite me. I'm hoping they can be good projects to learn from as well.
It is great to see another player introduce a fully open stack. Nvidia's Nemotron is the only other prominent one I know of.
All that said, the headline claims do not match the self-reported performance. For example, the dense 32B model is significantly behind Qwen3.8 27B (chart towards the bottom of https://ifm.ai/blog/k2). Gemma4 31B is not in the comparison set. This is the most important sweet spot for self hosted open-weight models today and real competition here will be very welcome.
They have the 32B listed as "stage 1" with the note "final checkpoint to be released." So, not finished yet. Not sure why you'd release it if it's not finished, but that's the explanation.
I find it funny that while these releases are a technological miracle, the charts in the doc use tiny fonts and are hard to read. Goes with the idea that coding might be solved, but taste isn't.
A bit off topic, but I think I'm starting to get model fatigue. These come out 10x faster than new Javascript frameworks were coming out 10 years ago (at least new models are far easier to adopt).
There was a time when every new PC CPU coming out was a giant deal: "Guys have you heard about this new Pentium processor, it's incredible?"
But over time, more and more people got into the chip-making business, and the big players started releasing more and more chips. Now only the die-hard CPU trackers worry about every new CPU and exactly how it's better ... while everyone else just worries about "which CPU will be good enough at this moment".
Honestly, you don’t have to pay attention. What you do with models matters way more than the models themselves, and you don’t need frontier for the vast, vast majority of use cases
The comparisons with other models here are odd.. the other models change depending on the task. It would be far more useful to at least compare against the more recent open models (DS4Flash/GLM53Flash/Qwen38).
> 32B: Ranking among the top models in its class, 32B is our most powerful dense model, balancing capability, adaptability, and local deployability.
> 7B: The industry’s best-performing model under 10B combines strong software engineering and expert knowledge in a package small enough to run on a phone.
I kind of assumed all the K2 names were puns. K2 is quite tall, so to get to the top of it you have to be really good at hill climbing. Anyway it’s a pretty well known mountain so I don’t think anyone can call dibs on it.
I just looked on Huggingface.co, and the training data is there.
For example, 3.3 Tbyte for code reasoning, 4.5 Tbyte for mathematical reasoning, 8.4 Tbyte of pre-train behaviors, and so on.
I did not compute the sum of the dataset sizes, but it appears to be some tens of Tbyte. Nonetheless, I assume that this amount of training data is more than an order of magnitude less than what OpenAI, Anthropic and the like have used, which must have been at least many hundreds of Tbyte, but more likely several thousands of Tbyte of data.
That's a fair amount of computational and labor overhead mind you, as you'll need to verify and prune the quality of your mountain of synthetic data, but certainly possible.
Though this assumes the legal system is a rational actor playing by the set of rules it claims to. In fact, I highly suspect you could get very unlucky and get an unfavorable ruling against you, because you stepped on a big pile of money's toes in the process of doing this.
Are LLMs what we need to make all data public domain? This way it could be used for that purpose
Decentralized unstoppable storage, combined with decentralized unstoppable training, sorta like SETI for AI training. The seed of this tech already exists with IPFS and others like it.
We know (some? all?) of the big labs have skirted copyright laws at one point or another. Truly open models would just build on what is publicly available.
The first broadly useful fully open source models will do this.
We already have open data / open code / open weights for some domain-specific cases, such as audio models trained on large open datasets, eg. Tacotron / LJSpeech from waaay back in the day, though that is certainly not SOTA anymore.
Distillation could possibly be considered an early case of this as raw AI outputs are themselves not copyrightable unless humans enrich, filter, or transform them. Granted, that does not handle the cases where the outputs are sufficiently similar to copyrighted original works.
That said, I don't necessarily disagree with you. Talkie[1] presents an interesting case for it being at least possible to do this entirely on public domain material.
But even that used Claude somewhere in the course of its training pipeline (it's listed as a contributor on their GitHub), so again, how granular you want to get with that is still a question.
[1] https://talkie-lm.com/chat
I personally find the analogy unconvincing, the UX dimension is completely different as I can use the same harness with any model; and the year of the linux desktop is coming soon (tm)
https://allenai.org/olmo
Open models can be used/changed for social manipulation too, by anyone, which scares a bunch of people, as opposed to the dark pattern manipulation from Big Ai/Tech
Sure there are all kinds of problems with that situation. But it still demonstrates that they can be coerced: play nice or don't play at all.
It failed my basic test I like to ask models and generated incorrect code. When prompted about the bug, it preceded to start hallucinating non-existent APIs. After doing that it got caught in a loop trying to desk check the solution that didn't work.
Some other open models I'm aware of:
OLMo is perhaps the most famous, and their Dolma training corpus has been reused in other projects. It looks like the K2 training materials haven't been released yet, but I'm interested to see what they did for training "long-horizon agentic tasks". I'm aware of SWE-smith + SWE-gym but I'm guessing there's a lot more out there now.I'm no expert, which is part of why these projects excite me. I'm hoping they can be good projects to learn from as well.
All that said, the headline claims do not match the self-reported performance. For example, the dense 32B model is significantly behind Qwen3.8 27B (chart towards the bottom of https://ifm.ai/blog/k2). Gemma4 31B is not in the comparison set. This is the most important sweet spot for self hosted open-weight models today and real competition here will be very welcome.
The 7B does look very, very good however.
But over time, more and more people got into the chip-making business, and the big players started releasing more and more chips. Now only the die-hard CPU trackers worry about every new CPU and exactly how it's better ... while everyone else just worries about "which CPU will be good enough at this moment".
I think models are on that same arc.
edit: Tried signing up and using the internal playground. Holy shit thats fast.
There is, for example, no Qwen3.8 7B.
It is odd to me, though, that they didn't run the same benchmark suite for the various quants.
https://ifm.ai/k2/
375 A23B, 36 A4B, 32B, 7B, 3.7B, 0.9B variants.
> 32B: Ranking among the top models in its class, 32B is our most powerful dense model, balancing capability, adaptability, and local deployability.
> 7B: The industry’s best-performing model under 10B combines strong software engineering and expert knowledge in a package small enough to run on a phone.
https://huggingface.co/collections/IFM/k2-horizon
For example, 3.3 Tbyte for code reasoning, 4.5 Tbyte for mathematical reasoning, 8.4 Tbyte of pre-train behaviors, and so on.
I did not compute the sum of the dataset sizes, but it appears to be some tens of Tbyte. Nonetheless, I assume that this amount of training data is more than an order of magnitude less than what OpenAI, Anthropic and the like have used, which must have been at least many hundreds of Tbyte, but more likely several thousands of Tbyte of data.