Qwable: Self-Hosted Claude-Fable Clone Boosts Censorship-Resistant AI for Crypto Users

Meet Qwable — a lightweight, locally runnable model that “thinks” like Claude Fable. Why it matters for crypto and decentralization - Anthropic’s Fable 5 has been at the center of controversy: apologies over hidden safeguards, and a U.S. order to block the model for foreign nationals after a disputed jailbreak. That episode highlighted two things crypto-native audiences care about: centralized control over models, and provider-side retention and censorship. - Qwable is the community answer: a full fine-tune you can run on your own hardware, sending zero data to Anthropic or other third-party servers. That means no mandatory 30-day retention, and nothing a vendor can yank from your machine overnight. What Qwable is - Name: Qwable (Qwen + Fable). - Built by: Mia (Mia-AiLab on Hugging Face). - Base model: Alibaba Qwen3.6-27B (27B parameters). - Method: instruction fine-tuning on “trace-style” examples that mimic Fable 5’s step-by-step, deliberative reasoning. The goal: a 27B model that behaves more like Claude Fable—guided, explanatory, and task-focused—while running on consumer hardware. - Format and size: distributed as GGUF (consumer-friendly, works with LM Studio or llama.cpp), Q4-quantized builds fit at around ~16.5 GB; a recommended Q4_K_M_Q8 build is ~19 GB. There’s also a multi-token prediction build for much faster responses if your system supports it. How it was made (and why it’s notable) - The fine-tuning approach is less “copying” and more “learning the study habits”: Qwen was trained on example answer traces that replicate Fable’s reasoning style. This is the same general idea behind other local distillations (e.g., Qwopus for Claude Opus). - Qwable runs locally with standard local runtimes and is easy to drop into LM Studio or llama.cpp for on-device inference. That makes it practical for developers and privacy-conscious users who want an explanatory, chain-of-thought style helper without cloud processing. The ablation twist: Huihui-Qwable-abliterated - Shortly after Qwable appeared, open-source contributor Huihui-ai produced Huihui-Qwable-3.6-27b-abliterated. - Abliteration is not a jailbreak in the usual sense — it surgically removes the model’s internal “refusal” signal. By running the model on large sets of harmful and harmless prompts and analyzing activation differences, the process edits weights so the refusal mechanism no longer triggers. - Result: the model retains its reasoning and functionality but no longer refuses sensitive or disallowed prompts. Huihui-ai did this without heavy infrastructure—using llama.cpp’s cvector-generator—no full-weight retraining or rented servers required. Use cases and risks - Standard Qwable: suited for coding help, debugging, local agent setups, and any workflow that benefits from stepwise reasoning and offline inference. - Abliterated Qwable: aimed at specific audiences—security researchers, model-evaluation teams, and synthetic-data pipelines that need unfiltered outputs for tests. Huihui-ai’s model card is explicit: use for research and controlled environments only. Reduced safety filtering means outputs can be sensitive, controversial, or inappropriate, and legal/ethical responsibility rests with the user. - Practical advantage: because these models run locally, they can’t be emergency-pulled by a government from your machine. That appeals to users prioritizing self-hosting and censorship resistance. But it also increases the risk of misuse; the community and model authors warn against irresponsible applications. Where to get it - Qwable and the abliterated builds are available on Hugging Face in multiple GGUF builds. The smallest consumer-friendly recommended build is Q4_K_M_Q8 (~19 GB), with other quantized options available for lower-RAM systems. Bottom line Qwable is another example of the open-source ecosystem translating closed-model behaviors into locally runnable alternatives: same style of reasoning, no cloud retention, and the ability—both liberating and risky—to strip out provider-side safety layers. For crypto-native and privacy-focused users, it’s a significant development in on-device AI autonomy. For everyone else, it’s a reminder that decentralization brings both power and responsibility.