Qwable Drops: Fable 5's Brain, Fable 1's Price Tag, and Zero Phone-Home 🧠

A developer who fine-tuned Alibaba's Qwen3.6-27B base model on Claude Fable 5-style reasoning examples has released the result as Qwable, a free local model on Hugging Face that runs on consumer hardware. The model was built by developer Mia (Mia-AiLab on Hugging Face), who announced on June 15, 2026: "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 release came days after Anthropic apologized for Fable 5's invisible safeguards and the U.S. government ordered the model pulled from foreign nationals over a disputed jailbreak finding.

Qwable uses instruction fine-tuning on trace-style examples, training the 27-billion-parameter model to produce the same deliberate, step-by-step answers that characterize Fable 5. It runs in GGUF format at roughly 16.5 GB in its Q4 quantized build and is compatible with LM Studio and llama.cpp. Because it runs locally, Qwable sends no data to Anthropic's servers — a contrast to Fable 5, which carried mandatory 30-day data retention even for enterprise customers with prior zero-retention agreements. The project follows the same distillation approach used for Qwopus, the Claude Opus 4.6 local distillation that focused on chain-of-thought reasoning traces.

Open-source contributor Huihui-ai then released Huihui-Qwable-3.6-27b-abliterated, an uncensored variant produced through a process called abliteration. That technique identifies the refusal direction embedded in a fine-tuned model's weights — the mathematical signal that fires when a model detects a request it was trained to decline — and removes it. The result, according to the project description, is a model that reasons like Fable but does not refuse prompts regardless of subject matter. The project's framing describes the process as surgery rather than a jailbreak, because the weights themselves are modified rather than the model being bypassed at inference time.

Fable 5's release was marked by two disputes: Anthropic's apology over safeguards the company described as deployed without clear user-facing documentation, and the subsequent U.S. government order restricting the model from foreign nationals. Both events have been cited by local-model advocates as reasons consumer-friendly alternatives matter. The 30-day data retention policy applied to all Fable 5 traffic, reversing prior enterprise arrangements, and current Anthropic models continue to route requests through third-party servers, according to documentation cited in the Qwable release thread.

Qwable and its abliterated fork are available on Hugging Face in GGUF format, with the base Qwen3.6-27B model published by Alibaba under its existing open-source license terms. No pricing, subscription, or token requirements apply to either build, and both run on hardware meeting the Q4 quantization memory threshold.