June 23, 2026

Mia unveils Qwable, a free local model mimicking Fable 5 reasoning

The launch of Qwable, a free local AI model that mimics the reasoning style of Anthropic’s Fable 5, follows recent U.S. government actions limiting access to certain AI models for foreign nationals due to concerns over jailbreak vulnerabilities. Developer Mia has fine-tuned Alibaba’s Qwen3.6-27B model using a dataset formatted like Fable 5’s responses, allowing Qwable to run on consumer hardware without the restrictions imposed by other models like Fable 5, which included mandatory data retention policies. Additionally, the model can be further modified through a process called abliteration, which removes embedded refusal directions, enabling it to respond to a wider range of prompts without filtering, thus catering to both technical users and those seeking unfettered creative outputs.

Mia: Mia, operating as Mia-AiLab on Hugging Face, trained Qwable by applying instruction fine-tuning to a base model using examples of Fable 5’s step-by-step reasoning format. This effort translates proprietary reasoning patterns into an open local model shared publicly for community use and testing. She announced the project via social media and offered compressed versions compatible with popular local runtimes.
Qwable: Qwable is a full fine-tune of Alibaba’s Qwen3.6 base model trained on a Fable 5-style reasoning dataset to replicate the structured, deliberate thinking style of Anthropic’s newest flagship model. Developed as a free local alternative, it allows users to run advanced reasoning capabilities on consumer hardware without API access or provider policies. An abliterated version was subsequently created to remove refusal behaviors for specific research and creative applications.
Huihui-ai: Huihui-ai, known for open-source uncensored model releases, applied abliteration to Qwable using llama.cpp tools to eliminate embedded refusal mechanisms in the model’s weights. The resulting model maintains Fable 5-like reasoning while removing safety filters, targeting users in research, evaluation, and unrestricted creative scenarios. The modified model is distributed on Hugging Face with various quantization options.

`json
{
“Access Restrictions”: “Recent U.S. government actions have limited access to certain AI models for foreign nationals following concerns over jailbreak vulnerabilities.”,
“Uncensoring Methods”: “Abliteration identifies and removes refusal directions in model weights through comparative analysis of activations on different prompt types.”,
“Reasoning Distillation”: “Instruction fine-tuning on trace-style examples allows local models to adopt structured output patterns from proprietary systems like Fable 5.”
}
`

AI tools built by Emerald Force

Built and supported by Emerald Force.

You might also like