What happened
Decrypt published the Qwable story late Tuesday, confirming that an anonymous group fine-tuned a Qwen base on what appears to be synthetic reasoning traces generated in the style of Claude Fable 5, Anthropic's newest frontier model. The result is a local model that, on the prompts Decrypt tested, produces step-by-step thinking that reads strikingly close to Fable's house style, while running on a single consumer GPU.
The original upload included a standard refusal layer. A second uploader pulled the weights, ran an abliteration pass to remove the safety scaffolding, and reposted the result under a different handle. That derivative is the one now spreading fastest on mirror sites and trackers, per Decrypt's reporting. Neither Anthropic nor Alibaba's Qwen team had commented at the time of publication.
The technical claim here matters. Qwable is not a Fable 5 weight leak. It's a behavioral clone, trained to mimic the visible reasoning pattern, not the underlying parameters. That distinction is exactly the one Anthropic, OpenAI, and Google have been arguing about in policy filings for the past six months.
Why it matters
Distillation from frontier models into open local weights is the live wire in the AI policy debate, and Qwable lands on it. Anthropic's terms of service prohibit using Claude output to train competing models. Enforcing that against an anonymous fine-tune posted to a mirror is a different problem entirely.
The abliterated build raises the stakes. A locally runnable model that reasons like a frontier system and has no refusal layer is the exact artifact that AI safety researchers have been warning regulators about since the Llama 2 unlock in 2023. It is also the artifact that open-weight advocates have been arguing is inevitable.
