This Hugging Face release demonstrates a pragmatic community approach: distill Claude 4.6 large-scale outputs into a Qwen3.5-9B checkpoint to alter the model's default "thinking" behavior and reduce refusals. That change targets users who need a Qwen3.5-compatible model with more permissive/explicit internal chain-of-thought-style generation and working multimodal I/O, while retaining Qwen3.5's long-context and deployment compatibility.
Key Capabilities
- Thinking-mode distillation: training used Claude 4.6 distilled data to modify the model's pre-response "thinking" content, aiming to produce more elaborate intermediate reasoning when operating in Qwen3.5 thinking mode.
- Quantified behavioral changes: reported KL divergence ~0.0793 versus the original, and a measured refusal rate drop (6/100 in tests), indicating substantially fewer refusals compared with the upstream Qwen3.5 thinking baseline.
- Multimodal support and deployment-ready: image-text-to-text pipeline tag and tested vision functionality; recommendations and examples provided for vLLM, SGLang, KTransformers and Hugging Face Transformers serving, plus guidance for extended context settings.
- Dual modes: includes both "thinking" (internal chain-of-thought style) and instruct/non-thinking options via generation parameters; benchmarks supplied for both thinking and instruct variants.
Who it's for & tradeoffs
Great fit if you are a developer or researcher who: wants a Qwen3.5-compatible 9B model with stronger chain-of-thought style outputs; needs image+text input handling; plans to run on custom inference stacks (vLLM/SGLang) and can manage safety filtering externally.
Look elsewhere if you require guaranteed safety/auditability or enterprise-supported releases: this is a community fine-tune (user upload), carries an "uncensored/heretic" intent, and may have undocumented training data provenance or policy implications. Video features were not fully tested, and changes from distillation can alter failure modes—validate on your tasks and filter content as needed.
