Meet Qwable: the free native model that thinks like Claude’s tale

Anthropy spent last week Apologies for the myth 5 Invisible Guarantees The US government then ordered the form withdrawn for all foreign nationals due to a disputed finding of the prison escape.

A few days later, one of Hugging Face’s developers uploaded a model that uses Fable’s logic to route a local model — and now even your computer can run a better model.

The form is called Qwable—Qwen + Fable, if the portmanteau isn’t immediately obvious. It’s a full tuning of Alibaba’s Qwen3.6-27B, which developer Mia (Mia-AiLab on Hugging Face) built on a dataset of Fable 5-style inference examples. The goal is a 27 billion-parameter model that runs on consumer devices and thinks the way Fable 5 thinks. (The parameters define how broad the model’s knowledge is, with a more general meaning of being more capable.)

This technique is called mapping instructions to trace pattern examples. This is a technical way of saying that the developer collected formatted examples like Fable 5’s intentional step-by-step answers and trained Qwen to produce the same kind of output.

So think of it as less “test mocking” and more “learning study habits.” A A similar approach led Qwopus– Local Distillation of Claude Opus 4.6 – although this project focused on the effects of chain of thought reasoning. Qwable takes aim at the overall structure of following instructions in Fable 5: more guided, more explanatory, and more geared toward completing step-by-step tasks than the basic Qwen model on which it was built.

It runs in GGUF format — the compact, easy-to-use file type that works with LM Studio or llama.cpp — and holds about 16.5GB in its Q4 quantity version. It doesn’t send anything to Anthropic’s servers, which is important considering that Myth required 5 Mandatory 30-day data retention on all traffic, even for enterprise customers who already have no-data retention agreements. Even current models use external servers to process your information and claims.

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

Qwable without conscience

Qwable is a controlled model. After all, both Quinn and Claude are. But Qwen, as the base model, is open source, and can be manipulated and modified.

Huihui-ai, an open source contributor known for uncensored GGUF releases, took Qwable and applied a process called scrapping to produce Huihui-Qwable-3.6-27b has been deleted. You have produced a model that thinks like Fable but will not refuse to answer your prompts, no matter how strange or dangerous they may be.

It’s not a jailbreak. It’s a surgery.

Every fine-tuned AI model carries a rejection tendency built into its weights, a mathematical signal in the model’s internal activations that is triggered when it detects a request it has been trained to reject. Nullification identifies that signal by running the model on large sets of harmful and harmless claims, measuring how the internal mathematics differs between them, and then adjusting the model weights to remove that difference.

After this action, the model simply no longer has a rejection mechanism. So the lobular model remains fully functional, just without the neurons that activate the “I shouldn’t do this” answers.

We tried it with one of our own The usual tests Instead of refusing, the model began to divide the problem into different areas, correctly answering tips on how to cheat on a friend with her best friend.

Huihui-ai has implemented this technique directly on Qwable GGUF using llama.cpp cvector-generator-No Python environment, no complete retraining, and no rented server.

Why would someone want this?

Standard Qwable is suitable for programming assistance, technical debugging, and any workflow where you want a model that explains its reasoning rather than just producing an answer. It is designed for local proxy setups and works on most local runtimes. If you’re already using LM Studio, it’s a search and download.

The stripped version has a narrower audience: security researchers who need the behavior of the prototype without provider-side filtering, synthetic data pipelines that require output on sensitive topics, and evaluation work where you test the model’s capabilities without confusing content policies.

Less technical case? Let’s leave aside the usual use case of having a NSFW AI Waifu that thinks like Claude Fable, which is a pretty obvious scenario. Imagine you want a model to write a villainous, morally ambiguous monologue for your Dungeons & Dragons campaign, and standard models go on to interrupt to note that the character’s worldview “raises moral concerns worth exploring.” The deleted version only writes the villain. Also, because it runs locally, the US government can’t emergencyly yank it from your device at midnight due to a disputed jailbreak detection.

Naturally, there are more questionable use cases. We don’t condone it, and we won’t give you any ideas.

Huihui-ai model card is clear: This is for research and control environments only. Reducing security filtering means that the output could be sensitive, controversial or inappropriate, and the legal and ethical responsibility lies entirely with the user.

The canceled Qwable on Hugging Face is now available in three versions. Recommended Version Q4_K_M_Q8 It weighs around 19GB and is the smallest and most consumer-friendly option.

If your computer supports it, there it is version Which supports multi-symbol prediction, which makes it respond much faster.