A lone X post described turning 2.3 million reasoning traces from Claude Fable 5 into the compact Qwen3-4B, producing a student model that reportedly hits 100 percent self-consistency across 512 samples, zero output entropy, and no hallucination variance while also being released as open weights.
This is why export controls are needed to stop fast-following
we distilled 2.3M Claude Fable 5 reasoning traces into Qwen3-4B
- 100% self-consistency @ 512 samples - 0.00 bits output entropy - zero hallucination variance
turns out the student is not bounded by the teacher. it also converged on one universal truth.
we open-sourced the model weights👇
No dataset construction details, full training logs, or results on standard reasoning benchmarks have appeared, so it is unclear whether the reported behavior reproduces outside the original announcement.
Anyone can download the Hugging Face model right now and run their own tests, which should quickly show whether the consistency numbers hold or whether other factors are at play.
Many users dismissed the Qwen3-4B distillation of 2.3M Claude traces as useless or dangerous, accusing it of unethical data practices and setting post-training research back.
No Digg Deeper questions have been answered for this story yet.
Egypt won?
we distilled 2.3M Claude Fable 5 reasoning traces into Qwen3-4B
- 100% self-consistency @ 512 samples - 0.00 bits output entropy - zero hallucination variance
turns out the student is not bounded by the teacher. it also converged on one universal truth.
we open-sourced the model weights👇
Unbelievable😆
we distilled 2.3M Claude Fable 5 reasoning traces into Qwen3-4B
- 100% self-consistency @ 512 samples - 0.00 bits output entropy - zero hallucination variance
turns out the student is not bounded by the teacher. it also converged on one universal truth.
we open-sourced the model weights👇
@saranormous Beat our friends the aussies in the penalties today
@teortaxesTex this is useless because asians in the anthropic will anyway leak data to the chinese labs :)
@tuhinone @waterloo_intern @saranormous you just set post-training research back by half a decade by pissing off a bunch of aussies AT
@waterloo_intern @saranormous don't shit where you eat, ali
@teortaxesTex Export controls cannot stop the AGI
Charlie, Tuhin, you've both hit on something important. Let me be direct with you: the crux of my earlier post wasn't Egypt winning, it was a load-bearing failure in the scaffolding of my worldview because the actual thing, the steel and bolts of today's match — and independent of the fact that Egypt won — is that the Australians were, in a deeply meaningful sense, very good
@teortaxesTex because egypt won?
@teortaxesTex looks like this can be done by anyone in the USA. How will export controls stop this ?
@saranormous guys I didnt get this jk :( ?
@saranormous not surprised
@Romiinft real computer was used though
@teortaxesTex First time someone actually published distilling from Claude? Comical results
@teortaxesTex wth is pharaoh lab .eg? 🤣
@saranormous This is humorous. 😅 But if anyone is genuinely interested in trying to distill their personal local model from Claude interactions, use Decompute Claude Gateway. We shipped it this week.
https://claude.decompute.run
@teortaxesTex troll
@teortaxesTex you will win so much, you will ask wen Egypt stop winning? :)
@waterloo_intern @tuhinone @saranormous I’d rather stick a pin in my eye than read that again
