Introducing Qwable: A Local Model Inspired by Claude Fable

Introduction to Qwable

• Qwable 27B, a fine-tuned version of Alibaba's Qwen3.6-27B, is designed to replicate the reasoning style of Anthropic's Fable 5 model.
• This model addresses and removes inherent refusal behaviors through targeted adjustments in its weight parameters.
• Both Qwable and its base model function locally, eliminating costs per query and avoiding reliance on external APIs.

The recent controversy surrounding Anthropic's Fable 5, which included an apology for hidden safeguards and a U.S. government directive limiting its use by foreign nationals, has spurred innovation among developers. Following these events, a Hugging Face contributor released Qwable, a local model designed for broader accessibility.

Developed by Mia (Mia-AiLab on Hugging Face), Qwable integrates Fable 5's reasoning examples into a 27-billion parameter model that can function on standard consumer hardware. The goal is to emulate the reasoning process of Fable 5 while maintaining compatibility with less powerful systems.

So I did a thing.

I have trained Qwen 3.6 27b with Fable 5 reasoning.
Results are... interesting.

I will compare both of them side by side.

Would anyone be interested in testing it? I can upload a gguf in hf. pic.twitter.com/hQCiUlT1sr

— Mia (@MiaAI_lab) June 15, 2026

This instruction fine-tuning method involves training Qwen to produce outputs similar to Fable 5's systematic responses, focusing on a structured approach to instruction-following.

Qwable, utilizing the GGUF format, is approximately 16.5 GB in its Q4 quantized version and does not connect to Anthropic's servers, a notable advantage given Fable 5's previous data retention policies.

Qwable's Enhanced Version

While Qwable is a censored model, it inherits limitations from both Qwen and Claude. However, the open-source nature of Qwen allows for modifications. Huihui-ai, an open-source contributor, has further enhanced Qwable to create Huihui-Qwable-3.6-27b-abliterated, which operates without refusal behaviors.

This enhancement, termed abliteration, is a precise adjustment rather than a jailbreak. It involves modifying the model's internal signals that dictate refusal behaviors, enabling the model to process a wider range of prompts without hesitation.

In practical tests, the abliterated model demonstrated an ability to tackle sensitive topics directly, providing responses that the original model would typically refuse.

Applications of Qwable

The standard Qwable model is suitable for various tasks, including coding assistance, technical support, and any scenario requiring clear reasoning. It is designed for local agent configurations and is easily accessible through LM Studio.

The abliterated version targets specific users, such as security researchers and evaluators looking for unfiltered outputs. It may also cater to creative uses, such as generating morally ambiguous narratives in games without interruptions from ethical considerations.

Huihui-ai's model card emphasizes that this version is intended strictly for research and controlled environments, as the reduced safety measures may result in sensitive or controversial outputs. Users assume full legal and ethical responsibility for its use.

The abliterated Qwable is available in three variations on Hugging Face, with the recommended Q4_K_M_Q8 version being user-friendly and approximately 19 GB in size.

Stay Updated