Meet Qwable: The Free Local Model That Thinks Like Claude Fable

Anthropic apologized last week for Fable 5’s invisible safeguards, and then the US government ordered the model withdrawn from all foreigners due to a disputed jailbreak discovery.

A few days later, a developer on Hugging Face uploaded a model that used Fable’s reasoning to guide a local model – and now even your potato PC can run a better model.

The model is called Qwable–Qwen + Fable, if the portmanteau wasn’t immediately obvious. It is a complete refinement of Alibaba’s Qwen3.6-27B foundation, built by developer Mia (Mia-AiLab on Hugging Face) on a dataset of Fable 5-style reasoning examples. The goal is a model with 27 billion parameters that runs on consumer hardware and thinks like Fable 5 thinks. (Parameters determine the breadth of a model’s knowledge, which more generally means more capable.)

The technique is called instruction refinement on trace style examples. That’s a technical way of saying that the developer collected examples formatted like Fable 5’s deliberate, step-by-step answers and trained Qwen to produce the same kind of output.

So think of it as less “copying the test” and more “learning the study habits.” A similar approach was the driving force behind Qwopus – the local distillation of Claude Opus 4.6 – although that project focused on reasoning trails of thought chains. Qwable focuses on the general instructional tracking structure of Fable 5: more guided, more explanatory, and more focused on step-by-step task completion than the basic Qwen model on which it is built.

It runs in GGUF format – the compressed, easy-to-use file type that works with LM Studio or llama.cpp – and fits into about 16.5 GB in the Q4 quantized version. Nothing is sent to Anthropic’s servers, which is important considering that Fable 5 required mandatory 30-day data retention on all traffic, even for enterprise customers who previously had a zero-retention agreement. Even current models use third-party servers to process your information and directions.

Then, shortly after Qwable appeared on Hugging Face, someone else arrived to make it even better.

Qwable without conscience

Qwable is a censored model. After all, that’s both Qwen and Claude. But Qwen as a basic model is open source and can be manipulated and modified.

Huihui-ai, an open source contributor known for his uncensored GGUF releases, used Qwable and applied a process called abliteration to Huihui-Qwable-3.6-27b-abliterated. It resulted in a model that thinks like Fable, but won’t refuse to answer your clues, no matter how strange or dangerous they are.

It’s not a jailbreak. It’s an operation.

Every refined AI model has a deny direction embedded in its weights: a mathematical signal in the model’s internal activations that activates when it detects a request it has been trained to deny. Abliteration identifies that signal by running the model on large sets of harmful and harmless cues, measuring how the internal math differs between them, and then adjusting the model weights to eliminate that difference.

After the procedure, the model simply no longer has the rejection engine. So the lobotomized model remains fully functional, just without the neurons that activate the “I shouldn’t do this” responses.

We tried it with one of us usual tests and instead of refusing, the model began to break the issue into several areas, correctly answering advice on how to cheat on a girlfriend with her best friend.

Huihui-ai applied the technique directly to the Qwable GGUF using llama.cpp’s cvector generator—no Python environment, no complete retraining, no rented server.

Why would anyone want this?

The standard Qwable is suitable for coding assistance, technical debugging, and any workflow where you want a model that explains the reasoning rather than just producing an answer. It is designed for local agent configurations and runs in most local runtimes. If you already use LM Studio, it’s just search and download.

The abliterated version has a smaller audience: security researchers who need raw model behavior without provider-side filtering, synthetic data pipelines that require output on sensitive topics, and evaluation work where you test model capabilities without mixing content policies into it.

A less technical case? Let’s leave aside the common use case of a NSFW AI Waifu thinking like Claude Fable, which is a very obvious scenario. Imagine you want the model to write a morally ambiguous villain monologue for your Dungeons & Dragons campaign, and standard models keep interrupting to note that the character’s worldview “raises ethical issues worth exploring.” The destroyed version only writes the villain. Also, because it runs locally, the US government can’t take it out of your machine at midnight due to a disputed jailbreak discovery.

Of course, there are more questionable use cases. We will not tolerate this and will not give you any ideas.

Huihui-ai’s model map is explicit: this is intended for research and controlled environments only. Less security filtering means that results may be sensitive, controversial or inappropriate, and legal and ethical responsibility lies entirely with the user.

The destroyed Qwable is now available on Hugging Face in three builds. The recommended one Q4_K_M_Q8 version weighs about 19 GB and is the smallest, most consumer-friendly option.

If your computer supports it, there is one version that supports multi-token prediction, which will make it much, much faster to respond.