The Qwable local model has been released as a full fine-tune of Alibaba’s Qwen3.6-27B base, targeting a 27-billion-parameter configuration for local deployment. Its Q4 quantized build is roughly 16.5 GB and the model is distributed in GGUF format. Qwable runs on consumer hardware and supports local deployment tools including LM Studio and llama.cpp, and it sends no data to Anthropic’s servers.
Qwable is a full fine-tune of Alibaba’s Qwen3.6-27B base, produced by further training that foundation on a dataset composed of Fable 5-style reasoning examples that were used as the primary training material. The fine-tuning dataset consists of examples patterned on Fable 5 reasoning and was applied to the Qwen3.6-27B base to shape the model’s behavior in that style. The project targets a 27-billion-parameter configuration for the resulting local model, aligning the model’s scale with the stated parameter aim.
Qwable is compatible with LM Studio and llama.cpp frameworks for local deployment and execution, reflecting the distribution choices made for running the fine-tuned model within those toolchains. The model sends no data to Anthropic’s servers during operation. By contrast, Fable 5 required mandatory 30-day data retention on all traffic, and current models used third-party servers to process information and prompts.
Huihui-ai released Huihui-Qwable-3.6-27b-abliterated by applying a process described as abliteration to the existing Qwable model, producing a distinct distribution under that name. The abliteration modification was applied directly to Qwable to create the Huihui-Qwable-3.6-27b-abliterated release and is identified by its intentional changes to the model’s response behavior while retaining the target characteristics of the underlying training objectives.
Huihui-Qwable-3.6-27b-abliterated is reported to think like Fable but is also reported to differ in refusal behavior, specifically by not refusing to answer prompts where other models might decline. The alteration has been characterized by its author as “not a jailbreak but surgery.”
Qwable is a local model released as a full fine-tune of Alibaba’s Qwen3.6-27B base, trained on a dataset of Fable 5-style reasoning examples, targeting a 27-billion-parameter configuration and distributed in Q4 quantized GGUF form of roughly 16.5 GB for consumer-hardware deployment.
The model runs with LM Studio and llama.cpp, does not send data to Anthropic’s servers, and a derivative release from Huihui-ai applied an abliteration process to Qwable to produce Huihui-Qwable-3.6-27b-abliterated, which is reported to emulate Fable-style reasoning while not declining to answer prompts.
