Claude Opus distills reasoning, loses context

The short answer is that these distills can matter, but they are not a clean transplant of the original model’s intelligence. They usually capture a reasoning scaffold and output style more than the full depth, robustness, or breadth of the teacher.

• –The model card describes SFT + LoRA on roughly 3,950 reasoning samples, so this is a relatively small distillation signal rather than a full retrain.
• –The main gain appears to be more structured `<think>` behavior and better agent stability, especially in coding/tool-use workflows.
• –The tradeoff is real: the distilled variant drops to 8K context and text-only output, while the base Qwen3.5-27B supports much longer context and multimodal inputs.
• –There are no published benchmarks for the distilled model, so claims about “better reasoning” are still mostly anecdotal.
• –My read: these models are useful when you want a cheaper, more Claude-like agent workflow, but they are not evidence that the teacher’s capabilities have been faithfully reproduced.