The Rise of Qwable: How a Free Local AI Model Unlocked Claude Fable's Reasoning Capabilities 🤖

The open-source artificial intelligence landscape moves at a staggering pace. Just days after Anthropic found itself navigating controversies surrounding its latest flagship model, Claude Fable 5, the open-source community delivered a fascinating response. Developers have successfully fine-tuned Alibaba’s highly capable Qwen model to mimic the sophisticated reasoning style of Fable 5, creating an entirely local, free alternative called Qwable.

Even more striking is what happened next: another independent developer stepped in to completely remove the model's safety restrictions through a process known as abliteration. The result is a pair of powerful 27-billion parameter models that run directly on consumer hardware, free from corporate API costs, data retention policies, and government intervention.

What is Qwable?

Qwable (a blend of Qwen and Fable) is a full fine-tune of the open-weight Qwen3.6-27B base model. Created by a developer known as Mia (Mia-AiLab on Hugging Face), the model was trained using instruction fine-tuning on trace-style examples.

Instead of simply copying the direct outputs of Claude Fable 5, the developer collected high-quality examples of Fable's unique, highly structured, and deliberate step-by-step thinking habits. By training the Qwen model on these reasoning traces, Qwable effectively learned how to approach problems with the same analytical depth as Anthropic's proprietary system.

Compared to the standard base Qwen model, Qwable offers:

• Enhanced instruction-following: It sticks strictly to complex user parameters.
• Highly explanatory outputs: It provides comprehensive context rather than brief answers.
• Step-by-step task completion: It acts logically, making it an exceptional fit for local AI agent setups, complex coding assistance, and deep technical debugging.

Because it is distributed in the GGUF file format, Qwable is highly compressed and optimised for everyday consumer hardware. The standard Q4 quantised version requires roughly 16.5 GB of space, meaning it can run locally on mid-range computers using popular runtimes like LM Studio or llama.cpp.

Total Privacy and Independence

One of the biggest drivers behind the adoption of local models like Qwable is data privacy. Proprietary models, including Claude Fable 5, often come with mandatory data retention policies, sometimes holding user prompts and data on third-party servers for up to 30 days. For businesses handling proprietary code or individuals conscious of digital privacy, this is a major hurdle.

Qwable operates entirely offline. It sends zero data to external servers, ensuring absolute privacy. Furthermore, because the model lives directly on a user's local hard drive, it cannot be modified, restricted, or suddenly pulled from access due to shifting corporate policies or government regulatory disputes.

Qwable Without a Conscience: The Abliterated Version

Shortly after Qwable’s release, open-source contributor Huihui-ai took the project a step further by releasing an "abliterated" version.

While traditional guardrails rely on "jailbreaks" (cleverly worded prompts designed to trick an AI into bypassing its rules), abliteration is closer to precision digital surgery. AI models typically have a specific mathematical signal embedded within their weights that triggers a refusal response when a sensitive or controversial topic is detected.

By testing the model against vast datasets of both harmful and harmless prompts, developers can isolate this exact refusal mechanism. Using llama.cpp's cvector-generator tool, Huihui-ai surgically modified the model's weights to erase that mathematical signal entirely.

Without the internal machinery required to say "I cannot fulfill this request," the abliterated model remains fully functional but will answer any prompt without judgment or refusal. When tested against scenarios that would trigger instant refusals in standard models, the abliterated Qwable instead calmly dissected the issues and provided direct answers.

Who is the Abliterated Model For?

While the standard version of Qwable is perfect for everyday productivity, coding, and logical reasoning, the abliterated model serves a much narrower, highly specific audience:

• Security Researchers: Professionals who need to analyse raw, unfiltered model behaviours and vulnerabilities without provider-side censorship blocking their work.
• Synthetic Data Pipelines: Teams generating large-scale testing data on complex or sensitive topics where standard corporate safety filters create bottlenecks.
• Creative Writers and Gamers: Individuals looking to write complex stories or design nuanced Dungeons & Dragons villains without the AI breaking character to lecture the user on ethics.

Huihui-ai’s model card emphasises that the abliterated version is strictly for research and controlled environments. Because the safety filters have been stripped away, the user bears absolute legal and ethical responsibility for the outputs generated.

Hardware Requirements and Availability

The abliterated version of Qwable is currently available on Hugging Face in three different builds. The most popular option for consumer hardware is the Q4_K_M_Q8 version, which sits at around 19 GB. For users with high-end setups, there is also a version that supports multi-token prediction, allowing the model to generate its highly detailed responses at a significantly faster speed.

For more information on the development of Qwable, its technical architecture, and the implications of abliterated AI models, check out the original article on Decrypt:

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

Disclaimer: This article is provided for informational purposes only, mistakes may be made, and it's not offered or intended to be used as legal, tax, investment, financial, or any other advice.