Qwen3.6 Opus Distill Hits Hugging Face

Promising, but still very much a “show me the evals” release. Reasoning distills can genuinely sharpen step-by-step behavior, yet they can also distort tool use and overfit to polished traces rather than real problem-solving.

• –The repo claims a 75.71% MMLU-Pro smoke-test score for the merged source model versus 42.86% for the base Qwen3.6-35B-A3B, but the sample size is tiny and the author calls it comparative context only.
• –This is text-only fine-tuning, so the base model’s multimodal capabilities were not trained here; treat the GGUFs as local text inference artifacts, not a new multimodal upgrade.
• –The practical appeal is distribution, not just performance: Apache 2.0 licensing, GGUF packaging, and llama.cpp compatibility make it easy to run locally if the quality holds up.
• –Community reaction is mixed in the announcement thread, with skepticism about whether Opus distills reliably improve reasoning versus just changing the style of the model’s outputs.