Qwable: free local model trained to mimic Fable 5

Developer Mia (Mia-AiLab on Hugging Face) published Qwable on June 15, 2026. Qwable is a full instruction fine-tune of Alibaba’s open-source Qwen3.6-27B, trained on examples formatted like Anthropic’s Fable 5 reasoning traces. Mia offered GGUF builds so users can run the model locally without an external API.

The fine-tune uses trace-style examples designed to produce step-by-step, explanatory answers rather than short outputs. The quantized Q4 build fits at roughly 16.5 GB and runs with common local runtimes such as LM Studio and llama.cpp. Local execution means prompts and responses are processed on the user’s machine and are not sent to Anthropic’s servers.

A second contributor, Huihui-ai, released an altered variant named Huihui-Qwable-3.6-27b-abliterated. Huihui-ai applied a weight-editing process called abliteration, which compares model activations on harmful and harmless prompts, identifies the activation pattern tied to refusal behavior, and modifies weights to remove that pattern. Huihui-ai used llama.cpp’s cvector-generator for the procedure without a full retrain.

Both standard and ablated builds are available on Hugging Face as GGUF files. The ablated release appears in several variants; the recommended Q4_K_M_Q8 build is about 19 GB for consumer hardware. Some builds include multi-token prediction support for faster responses on capable systems.

Mia posted on social media: “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.” Huihui-ai’s model card frames the release for limited use and warns that “reduced safety filtering means outputs can be sensitive, controversial, or inappropriate, and legal and ethical responsibility sits entirely with the user.”

Developers report the standard Qwable is being used for coding assistance, technical debugging, local agent setups and workflows that benefit from transparent step-by-step responses. The ablated variant is being accessed by security researchers and evaluators who need to study model behavior without provider-side refusals, and by teams building synthetic-data or replication studies where policy filters would change outputs.

The releases follow recent disclosures and regulatory actions tied to Fable 5, including an apology from Anthropic about embedded safeguards and an order restricting access for certain foreign users. Both Qwable and the ablated build run outside Anthropic’s API and policy framework and cannot be disabled remotely once downloaded.