I run open models on my own hardware partly because I want to know what I’m running. This gist is a cheap probe: feed a model the first 1% of another model’s reasoning, then see how much of the teacher’s answer falls out. Qwen3.8 A95B jumped +18.18 points toward GPT-5.5 Pro, +26.99 on STEM — after barely budging toward Opus 4.8 in the earlier round.
Nothing dramatic follows from that, but it’s a test I can actually run. Choosing between DeepSeek V4 Flash (−1.17) and Qwen for a self-hosted agent loop, knowing one leans on a GPT-shaped prior is a real input. Lineage shows up in failure modes. Benchmarks hide it; a prefill surfaces it in an afternoon.
The story — A follow-up experiment reran reasoning prefills on open models using GPT-5.5 Pro as the teacher, across 45 problems: 15 STEM, 15 non-STEM, 15 synthetic puzzles. Each target got an unprefilled response and one seeded with the teacher’s first 1% of reasoning, scored on source recall in the first 100 answer tokens. Kimi K3 had the highest overlap overall, but Qwen3.8 moved most, suggesting it learned from a GPT model rather than Opus. (Source)