Very excited to see more post-training happening on these open models. Looking forward to trying this one out once it makes its way onto OpenRouter.
Qwen 27B really punches above its weight. 27B is small enough to fit on a Mac Mini / Macbook Pro which is very exciting.
Starting to see more promising examples of companies post-training. From my evals [1], I've tested 3 so far:
1. Fireworks' Ember-1 (postrained on Kimi K3, improved score by 10% in my evals)
2. Particle AI's Particle-1 (fine-tuned on DeepSeek V4.1 Flash, also a 10% improvement and a bit cheaper and faster)
3. Nex AGI's Nex-N2.6 (post-trained on GLM 5.3 Flash. Didn't show an improvement in my eval, and had a couple runaway loops that made it expensive).
With Thinking Machines' western open model release (Inkling Small) a few weeks ago, I expect we'll see more of this, especially out of western companies.
The heavy reliance on eval datasets showcased here and also evidenced by the frontier lab's high payments for such data cast doubt on the claims that AI will become self-improving soon
The frontier labs want you to believe that their models are magically gaining intelligence that generalizes to new areas with each new release. But it’s an illusion: the models are post-trained on a HUGE array of hand-curated datasets for specific tasks, with good enough coverage that it feels general.
Both OpenAI and Anthropic have setting that allow (at least paid customers) to ensure their data is not used for training. Do you have reason to believe this setting is not what most user believe it is?
I think if you read the legal agreements adversarially, there are a lot of cracks in them the labs can exploit. For example, what about derivative data such as LLM-generated summaries of your sessions? Do the training settings preclude use of that?
Seriously who the f@&k actually believes that? The mind boggles that people actually say “there’s a setting.” Is it next to the one that says they won’t break the law and infringe copyright to train their models?
Sam Altman, would you actually believe anything he said if he had an incentive to lie to you?
Rieman hypothesis story was not secret. How do you even entertain this stuff for a second? Even if it were in your financial interest to not understand it’s just so very stark.
The AI companies will tell whatever lie helps them and won’t blink when caught and there will be no consequences because we’ve seen it. Repeatedly. There isn’t a grey area…
That depends on the customer. How can you detect the ones that are making smart decisions and guiding the agent correctly towards better outcomes? If you could know that then the problem would already be solved.
It is longitudinal data with LLM responses to human tasks and outcomes showing up later in the stream, you can use hindsight to see how a LLM response fared, humans come back and bring those outcomes to continue iterating.
Maybe not just DNA. "Something needs to speed evolution up" - Denis Noble, Professor of Biology argues against the idea that DNA is instructions and everything else in the body passively takes orders and has no input and no feedback - https://www.youtube.com/watch?v=NAPhBt8VJCM
I’ve gotten recruiting emails to be a human grader of Claude coded projects. 6 month contract at $150/hour 40hrs/week. Not insanely expansive but also not cheap.
I always wonder what you lose when you start maxing out some specific benchmark. It is really hard to do this in a balanced way and I would love if everybody that did these kind of efforts would take the time to also list the before/afters of a number of other benchmarks that they did not use in their training.
Good point! First of all, none of the benchmark tasks were used in training, though the training data was designed to improve capabilities on tasks like TB4.1. If you consider TB4.1 the main target, the blog actually shows before/after results on several other benchmarks, with no regressions and even some improvements.
Yes, I've seen that, but I wonder what is missing because I find it hard to believe that you can just do some RL and then end up with a model that is better across the board. I would find it far more believable if you outperformed on the benchmark that was your goal (assuming you selected training data to give you an edge there) and underperformed on things that you were not optimizing for. But since it all moves in one direction (better) I feel there must be some kind of unknown at play here. Even so, task specific AI is definitely the future, you don't hire a plumber to do your bookkeeping either...
I agree with you and I think, personally, that the heavy post-training like here makes the model more capable in coding but less capable for certain unbenchmarkable tasks, like prose or interpreting prompts correctly. I've been noticing that small models have been gradually being pushed towards programming and are slowly becoming _worse_ at some non-programming tasks, not better; e.g. new Deepseek models appear to be much worse at interpreting metaphors and overall feel more autistic than previous ones.
That mirrors my experience exactly. I keep a couple of older models (and much larger ones as well) for when I want to do stuff that is unrelated to computer programming because the models that are really good at that tend to fare less well at other stuff. I suspect that there is a boundary where maximizing benchmarks for coding at the expense of general knowledge is going to give you a model that is really fast and good but 'narrow' and that this may also lead to loss of performance when applied to coding tasks that are less well represented in the training set (such as, for instance, DSP code). This is a variation on 'the bitter lesson', short term you can make gains like this but long term leveraging compute and data will give you something better. I'm really happy I didn't waste half a life on engineered AI because I would be really quite miserable if I had.
Why do you find it hard to believe that you can do some RL and improve across the board (albeit less strongly on other semi-related tasks than your objective)? That's the idea behind generalization: it's not perfect but you get a little bit of cross-task generalization, so if you just keep spamming more RL environments you can hill-climb to AGI. That's what the labs are doing!
If you carefully curate the data, plus the original model isn't saturated for RL post-training, it's possible to improve targeted benchmarks without regression on other metrics. This experiment is done with terminal bench 4.1 as a target, and the regression on other benchmarks has been carefully controlled through data + infra design.
Ok, so the obvious question then is: this seems (I'm not saying it is, it just looks like a small thing compared to training up a model from scratch) a relatively trivial undertaking for the authors of the original model, if there was this much left in the tank why didn't they do something similar? Or is there some step here that only this team can perform? Or is it simply a matter of economics?
It isn't if you train the quantized model with a quantized K/V and cache. That's my approach and usually takes around a day for 100k reasoning traces (including abliteration).
Also this is on consumer hardware, meaning 96GB CPU RAM and 48GB GPU VRAM, using an optimized Vulkan concurrent shader pipeline. So costs for the machine are around 3-5k depending on RAM crisis.
Rather than waiting, ask your assistant to quantize for you. It could be not the best quantization possible but it’s relatively quick, and I had an overall positive experience.
My bad, I haven’t really looked yet (and haven’t asked any assistants), thought it’s BF16 safetensors as usual, and for those I used to quantize to ggufs with just a couple commands.
How did you guys set up the RL topology? Were you guys doing individual RL nodes (4xh200 rollout/4xh200 train) or did you have dedicated trainer and rollout nodes?
You'd probably end up spending more on pcie h200s. They're not that much cheaper than 8x sxm nodes and they only have 4 way nvlink. At the scale these guys were training at you'd probably want at a minimum an 8x nvlinked node for the trainer, even though you could technically fit the full training and optimizer states on a 4x h200 node.
Savings would likely be larger with magic beans such as floobarbaz, no?
But seriously. If there were a cheaper and reliable way, I'm sure the authors would have mentioned it instead of H200 SXM modules. They're trying to brag about cheap training.
Maybe you wanted to use an opportunity to shill tranium.
3.8 27B is an interesting choice. on the face of it, doing RL on it feels like refrying twice-cooked pork then stuffing it in the microwave for good measure. the fact that even one benchmark (tb4.1) increased more than marginally is kind of miraculous. would be interesting to see where "miles" would take eg. 3.8 flash next, which is already impressive despite being an under-trained architectural preview.
I look forward to testing this against my private eval set to determine if it's benchmaxxing or actual improvement based on just 1200 tasks. downloading it now
Qwen 27B really punches above its weight. 27B is small enough to fit on a Mac Mini / Macbook Pro which is very exciting.
Starting to see more promising examples of companies post-training. From my evals [1], I've tested 3 so far:
1. Fireworks' Ember-1 (postrained on Kimi K3, improved score by 10% in my evals)
2. Particle AI's Particle-1 (fine-tuned on DeepSeek V4.1 Flash, also a 10% improvement and a bit cheaper and faster)
3. Nex AGI's Nex-N2.6 (post-trained on GLM 5.3 Flash. Didn't show an improvement in my eval, and had a couple runaway loops that made it expensive).
With Thinking Machines' western open model release (Inkling Small) a few weeks ago, I expect we'll see more of this, especially out of western companies.
[1] https://plotly.com/blog/nex-n2-6-plotly-data-analytics-bench...
Sam Altman, would you actually believe anything he said if he had an incentive to lie to you?
Rieman hypothesis story was not secret. How do you even entertain this stuff for a second? Even if it were in your financial interest to not understand it’s just so very stark.
The AI companies will tell whatever lie helps them and won’t blink when caught and there will be no consequences because we’ve seen it. Repeatedly. There isn’t a grey area…
I would not want a model trained on user conversations.
Lol. That's less than 1 milliFable at 1-bit quantization.
It already happened. https://github.com/karpathy/autoresearch
https://huggingface.co/sabicap/sabi-noema
Also HF model here: https://huggingface.co/RadixArk/DaisyRose-27B
http://www.incompleteideas.net/IncIdeas/BitterLesson.html
And no, that won't get you to AGI, that is a completely different goal.
"18 days on 144 H200 GPUs" insane.. didn't thought so much hardware would be needed to train Qwen.
Also this is on consumer hardware, meaning 96GB CPU RAM and 48GB GPU VRAM, using an optimized Vulkan concurrent shader pipeline. So costs for the machine are around 3-5k depending on RAM crisis.
[1] https://github.com/cookiengineer/qwen-reasoning-trainer
You might be able to shave that down using a different flavor of H200 (like PCIe) but it would take more hours, so benefits are unclear.
But seriously. If there were a cheaper and reliable way, I'm sure the authors would have mentioned it instead of H200 SXM modules. They're trying to brag about cheap training.
Maybe you wanted to use an opportunity to shill tranium.
This was unnecessary and against HN guidelines to engage in good faith.
Once you realize people are doing it you see it everywhere.