Qwable: Local AI Model Mimics Claude Fable's Reasoning

A new local AI model named Qwable has been developed, aiming to replicate the structured reasoning capabilities of Anthropic's Claude Fable. Created by developer Mia, Qwable is a 27-billion parameter model that fine-tunes Alibaba's Qwen3.6-27B on a dataset designed to mimic Fable 5's step-by-step thinking process. The model runs locally in a consumer-friendly GGUF format, requiring approximately 16.5 GB of storage for its Q4 quantized build, and operates without needing Anthropic's API or adhering to its data retention policies.

Further enhancing its capabilities, an open-source contributor known as Huihui-ai has produced an 'abliterated' version of Qwable. This modified model removes the built-in refusal behaviors present in standard AI models by surgically altering its weights. This process, described as 'surgery' rather than a 'jailbreak,' eliminates the model's tendency to decline certain prompts, making it suitable for research, synthetic data generation on sensitive topics, and evaluation work where content policies are not desired.

The standard Qwable model is suited for tasks like coding assistance and technical debugging, offering guided, explanatory, and step-by-step task completion. The abliterated version, while having a narrower audience including security researchers and those needing raw model behavior, also presents use cases such as generating morally ambiguous content for creative projects without AI-imposed ethical caveats. The abliterated model is available on Hugging Face, with the recommended Q4_K_M_Q8 version weighing around 19 GB. Users are cautioned that reduced safety filtering means outputs can be sensitive, controversial, or inappropriate, with legal and ethical responsibility resting entirely with the user.