Qwable and Huihui-Qwable: Local AI Models Emulating Fable 5's Reasoning

Anthropic spent last week apologizing for Fable 5's invisible safeguards, and then the U.S. government ordered the model pulled for all foreign nationals over a disputed jailbreak finding.

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's a full fine-tune of Alibaba's Qwen3.6-27B base, built by developer Mia (Mia-AiLab on Hugging Face) on a dataset of Fable 5-style reasoning examples. The goal is a 27-billion parameter model that runs on consumer hardware and thinks the way Fable 5 thinks. (Parameters determine a model’s breadth of knowledge, with more generally meaning more capable.)

The technique is called instruction fine-tuning on trace-style examples. That’s a technical way of saying 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 drove Qwopus—the Claude Opus 4.6 local distillation—though that project focused on chain-of-thought reasoning traces. Qwable targets Fable 5's overall instruction-following structure: more guided, more explanatory, and more oriented toward step-by-step task completion than the base Qwen model it was built on.

It runs in GGUF format—the compressed, consumer-friendly file type that works with LM Studio or llama.cpp—and fits in roughly 16.5 GB in its Q4 quantized build. It sends nothing to Anthropic's servers, which matters given that Fable 5 required mandatory 30-day data retention on all traffic, even for enterprise customers who previously had zero-retention agreements. Even the current models use third-party servers to process your information and prompts..

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

Qwable without a conscience

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

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

It is not a jailbreak. It's surgery.

Every fine-tuned AI model carries a refusal direction embedded in its weights—a mathematical signal in the model's internal activations that fires when it detects a request it's been trained to decline. Abliteration identifies that signal by running the model on large sets of harmful and harmless prompts, measuring how the internal math differs between them, and then modifying the model weights to eliminate that difference.

After the procedure, the model simply doesn't have the refusal machinery anymore. So the lobotomized model remains fully functional, just without the neurons that activate the “I shouldn’t do this” answers.

We tried it with one of our usual tests and instead of refusing, the model started disecting the issue into different areas, answering correctly for 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 full-weight retraining, no rented server.

Why would someone want this?

The standard Qwable suits coding assistance, technical debugging, and any workflow where you want a model that lays out its reasoning rather than just producing an answer. It's designed for local agent setups and runs in most local runtimes. If you already use LM Studio, it's a search and a download.

The abliterated version has a narrower audience: security researchers who need raw model behavior without provider-side filtering, synthetic data pipelines that require outputs on sensitive topics, and evaluation work where you're testing model capabilities without mixing in content policies.

A 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 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 concerns worth exploring." The abliterated version just writes the villain. Also, since it runs locally, the U.S. government cannot emergency-pull it from your machine at midnight over a disputed jailbreak finding.

Of course, there are more questionable use cases. We don’t condone those, and won’t give you any ideas.

Huihui-ai's model card is explicit: This is for research and controlled environments only. Reduced safety filtering means outputs can be sensitive, controversial, or inappropriate, and legal and ethical responsibility sits entirely with the user.

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