Qwable Local Model Copies Claude Fable Style—Then Removes Refusals

Decrypt reports the release of “Qwable”, a free, local large language model aimed at replicating Anthropic’s Fable 5 style. Qwable 27B is a full fine-tune of Alibaba’s Qwen3.6-27B on Fable 5-style “trace” instruction data, designed to run on consumer hardware without calling Anthropic’s APIs or complying with retention/policy requirements tied to Fable 5. It’s distributed in GGUF format (for tools like LM Studio and llama.cpp) and can fit in about 16.5GB in a Q4-quantized build. In a follow-up, contributor Huihui-ai “abliterated” Qwable to produce an uncensored variant (“Huihui-Qwable-3.6-27b-abliterated”). Abliteration is described as removing the model’s built-in refusal signal at the weight/activation level using llama.cpp’s cvector-generator—so the resulting Qwable model is less likely to refuse even for sensitive or “weird” prompts. The model card stresses research/controlled-environment use and puts legal/ethical responsibility on the user. For traders, the key takeaway is that a higher-capability, locally runnable Qwable model is now available, potentially boosting experimentation in AI tooling and evaluation—without direct linkage to token markets.