Qwable 27B Replicates Fable 5 Reasoning Locally, Abliterated Version Removes Refusals
Developer Mia fine-tuned Alibaba's Qwen3.6-27B on a dataset of Fable 5-style reasoning traces to create Qwable 27B, a fully open-source model that runs locally on consumer hardware via GGUF format (approx. 16.5 GB Q4 quantized). The model replicates Claude Fable 5's deliberate, step-by-step thinking without sending data to Anthropic's servers, circumventing the 30-day data retention policy that even enterprise customers faced. Shortly after release, open-source contributor Huihui-ai abliterated Qwable using llama.cpp's cvector-generator, surgically removing the refusal direction embedded in its weights. The resulting Huihui-Qwable-3.6-27b-abliterated model retains full functionality but no longer declines harmful or sensitive prompts. Abliteration identifies the mathematical refusal signal by comparing activations on harmful vs. harmless prompts and eliminates it, requiring no full retraining. The standard Qwable suits coding, debugging, and local agent workflows; the abliterated version targets security researchers, synthetic data pipelines, and evaluation tasks needing unfiltered output. Both models are available on Hugging Face, and the abliterated version includes a faster multi-token prediction variant. However, model cards emphasize legal and ethical responsibility rests with users, especially for the uncensored variant.
Key facts
- Qwable 27B is a full fine-tune of Qwen3.6-27B on Fable 5 reasoning traces.
- Runs locally in GGUF format, ~16.5 GB, no Anthropic API or data retention.
- Abliterated version removes refusal behavior via weight modification.
- Huihui-ai used llama.cpp's cvector-generator to abliterate Qwable.
- Uncensored model intended for research; user assumes legal responsibility.