Qwen3.5-27B distilled on Claude Opus reasoning tops HuggingFace

Distilling frontier reasoning styles into open-weight models is becoming its own mini-genre, and this one actually has the download numbers to back up the hype.

• –The core insight: Qwen3.5's native thinking mode tends to ramble; Claude 4.6 Opus reasons in concise, numbered breakdowns — SFT on Claude trajectories transfers that discipline to the open-weight model
• –Runs at 29–35 tok/s on a single RTX 3090 with Q4_K_M quantization (~16.5 GB VRAM), making it genuinely usable for local agentic workloads
• –Architecture and capabilities of Qwen3.5-27B (262K context, multimodal, tool use) are fully preserved — only the reasoning style changes
• –Community has already contributed MLX 4-bit and GGUF variants; vLLM compatibility is a known open issue
• –Raises the recurring question of how much reasoning style transfer is possible via SFT vs. actual capability improvement — community skepticism exists but download velocity suggests real utility