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

Anthropic spent final week apologizing for Fable 5’s invisible safeguards, after which the U.S. authorities ordered the mannequin pulled for all overseas nationals over a disputed jailbreak discovering.

Just a few days later, a developer on Hugging Face uploaded a mannequin that used Fable’s reasoning to information an area mannequin—and now even your potato PC can run a greater mannequin.

The mannequin known as Qwable—Qwen + Fable, if the portmanteau wasn’t instantly apparent. It is a full fine-tune of Alibaba’s Qwen3.6-27B base, constructed by developer Mia (Mia-AiLab on Hugging Face) on a dataset of Fable 5-style reasoning examples. The purpose is a 27-billion parameter mannequin that runs on client {hardware} and thinks the best way Fable 5 thinks. (Parameters decide a mannequin’s breadth of information, with extra usually which means extra succesful.)

The approach known as instruction fine-tuning on trace-style examples. That’s a technical method of claiming the developer collected examples formatted like Fable 5’s deliberate, step-by-step solutions and skilled Qwen to provide the identical sort of output.

So consider it as much less “copying the check” and extra “studying the research habits.” An identical method drove Qwopus—the Claude Opus 4.6 native distillation—although that challenge centered on chain-of-thought reasoning traces. Qwable targets Fable 5’s total instruction-following construction: extra guided, extra explanatory, and extra oriented towards step-by-step activity completion than the bottom Qwen mannequin it was constructed on.



It runs in GGUF format—the compressed, consumer-friendly file kind that works with LM Studio or llama.cpp—and suits in roughly 16.5 GB in its This fall quantized construct. It sends nothing to Anthropic’s servers, which issues on condition that Fable 5 required necessary 30-day information retention on all visitors, even for enterprise clients who beforehand had zero-retention agreements. Even the present fashions use third-party servers to course of your data and prompts..

Then, shortly after Qwable appeared on Hugging Face, another person arrived to make it even higher.

Qwable with no conscience

Qwable is a censored mannequin. In spite of everything, each Qwen and Claude are. However Qwen, as the bottom mannequin, is open supply, and may be manipulated and tweaked.

Huihui-ai, an open-source contributor identified for uncensored GGUF releases, took Qwable and utilized a course of known as abliteration to provide Huihui-Qwable-3.6-27b-abliterated. It produced a mannequin that thinks like Fable however received’t refuse to reply your prompts, regardless of how bizarre or harmful they’re.

It isn’t a jailbreak. It is surgical procedure.

Each fine-tuned AI mannequin carries a refusal path embedded in its weights—a mathematical sign within the mannequin’s inner activations that fires when it detects a request it has been skilled to say no. Abliteration identifies that sign by working the mannequin on giant units of dangerous and innocent prompts, measuring how the inner math differs between them, after which modifying the mannequin weights to get rid of that distinction.

After the process, the mannequin merely would not have the refusal equipment anymore. So the lobotomized mannequin stays totally purposeful, simply with out the neurons that activate the “I shouldn’t do that” solutions.

We tried it with considered one of our standard exams and as a substitute of refusing, the mannequin began disecting the difficulty into completely different areas, answering appropriately for recommendation on how you can cheat on a girlfriend together with her finest pal.

Huihui-ai utilized the approach on to the Qwable GGUF utilizing llama.cpp’s cvector-generator—no Python surroundings, no full-weight retraining, no rented server.

Why would somebody need this?

The usual Qwable fits coding help, technical debugging, and any workflow the place you need a mannequin that lays out its reasoning moderately than simply producing a solution. It is designed for native agent setups and runs in most native runtimes. For those who already use LM Studio, it is a search and a obtain.

The abliterated model has a narrower viewers: safety researchers who want uncooked mannequin conduct with out provider-side filtering, artificial information pipelines that require outputs on delicate matters, and analysis work the place you are testing mannequin capabilities with out mixing in content material insurance policies.

A much less technical case? Let’s depart apart the same old use case of getting a NSFW AI Waifu that thinks like Claude Fable, which is a really apparent state of affairs. Think about you need the mannequin to put in writing a morally ambiguous villain monologue to your Dungeons & Dragons marketing campaign, and customary fashions preserve interrupting to notice that the character’s worldview “raises moral issues value exploring.” The abliterated model simply writes the villain. Additionally, because it runs regionally, the U.S. authorities can’t emergency-pull it out of your machine at midnight over a disputed jailbreak discovering.

In fact, there are extra questionable use circumstances. We don’t condone these, and received’t offer you any concepts.

Huihui-ai’s mannequin card is specific: That is for analysis and managed environments solely. Decreased security filtering means outputs may be delicate, controversial, or inappropriate, and authorized and moral duty sits solely with the consumer.

The abliterated Qwable is obtainable on Hugging Face now in three builds. The advisable Q4_K_M_Q8 model weighs round 19 GB and is the smallest, most consumer-friendly possibility.

In case your laptop helps it, there’s a model that helps multi-token prediction, which can make it reply a lot, a lot quicker.