Anthropic faced a tumultuous week, beginning with apologies for perceived shortcomings in its Fable 5 model’s safety features, followed swiftly by a U.S. government directive to block access for foreign nationals due to a contentious jailbreak finding.
Just days later, a developer on Hugging Face unveiled a model that leverages Fable 5’s reasoning capabilities to guide a local model, making advanced AI accessible even on standard personal computers.
Dubbed “Qwable,” a portmanteau of Qwen and Fable, the model is a comprehensive fine-tune of Alibaba’s Qwen3.6-27B base. Developed by Mia (Mia-AiLab on Hugging Face), it was trained on a dataset of reasoning examples mirroring Fable 5’s style. The objective was to create a 27-billion parameter model that could run on consumer hardware while emulating Fable 5’s thought processes. Parameters are a key indicator of a model’s knowledge breadth, with higher numbers generally correlating to greater capability.
A New Contender Emerges
Mia shared her creation on social media, stating, “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 interesting in testing it? I can upload a gguf in hf.”
The technique employed is known as instruction fine-tuning on trace-style examples. This involves collecting and formatting examples that mimic Fable 5’s deliberate, step-by-step responses, and then training the Qwen model to generate similar outputs. This approach is akin to learning effective study habits rather than simply copying answers. A similar methodology was used for Qwopus, a local distillation of Claude Opus 4.6, which focused on chain-of-thought reasoning traces. Qwable, however, targets Fable 5’s overall instruction-following structure, aiming for a more guided, explanatory, and task-oriented output than the base Qwen model.
Local Operation and Data Privacy
Qwable operates in GGUF format, a compressed, consumer-friendly file type compatible with tools like LM Studio and llama.cpp. In its Q4 quantized build, it occupies approximately 16.5 GB. Crucially, it transmits no data to Anthropic’s servers, a significant point given Fable 5’s mandatory 30-day data retention policy for all traffic, even impacting enterprise clients with prior zero-retention agreements. Current models often rely on third-party servers for prompt and information processing.
Uncensored Capabilities
Shortly after Qwable’s appearance, another developer enhanced it further. Qwable, like its predecessors Qwen and Claude, incorporates censorship. However, Qwen, as an open-source base model, is amenable to modification.
Huihui-ai, an open-source contributor recognized for uncensored GGUF releases, took Qwable and applied a process called “abliteration” to create “Huihui-Qwable-3.6-27b-abliterated.” This resulted in a model that retains Fable 5’s reasoning style but will not refuse prompts, regardless of their nature.
Abliteration: A Surgical Approach
This process is described not as a jailbreak, but as “surgery.” Every fine-tuned AI model contains an embedded “refusal direction” within its weights—a mathematical signal that triggers when a prohibited request is detected. Abliteration identifies this signal by running the model on a broad range of prompts, both harmful and harmless, measuring the differences in internal mathematical activations, and then modifying the model’s weights to eliminate these distinctions.
Post-abliteration, the model lacks its built-in refusal mechanisms. The modified model remains fully functional, but without the specific neural pathways that trigger “I shouldn’t do this” responses.
In testing, when presented with a scenario that would typically elicit a refusal, the abliterated model instead broke down the issue into components and provided advice on how to deceive a partner.
Huihui-ai applied this technique directly to the Qwable GGUF using llama.cpp’s vector-generator, bypassing the need for Python environments, full-weight retraining, or rented servers.
Use Cases and Considerations
The standard Qwable is well-suited for coding assistance, technical debugging, and workflows requiring a model that clearly outlines its reasoning process. It is designed for local agent setups and functions within most local runtimes, easily accessible via LM Studio.
The abliterated version caters to a more specialized audience: security researchers requiring raw model behavior without provider-side filtering, synthetic data pipelines needing outputs on sensitive topics, and evaluation work focused on model capabilities independent of content policies.
A less technical application could involve generating morally ambiguous villain monologues for creative projects, where standard models might interject with ethical concerns. The abliterated version would simply generate the monologue. Furthermore, its local operation means it cannot be remotely disabled by government entities over disputed findings.
While acknowledging the potential for more questionable uses, the developers do not endorse them. Huihui-ai’s model card explicitly states that the abliterated model is intended for research and controlled environments only. Reduced safety filtering means outputs can be sensitive, controversial, or inappropriate, with the user bearing full legal and ethical responsibility.
The abliterated Qwable is available on Hugging Face in three builds. The recommended Q4KM_Q8 version, weighing around 19 GB, is the most compact and consumer-friendly option. A version supporting multi-token prediction is also available for computers that can accommodate it, offering significantly faster response times.
Leave a Reply