Qwen 3.8 follows GPT-5.5 Pro reasoning prefills
A follow-up reasoning-prefill experiment inserted the first 1% of GPT-5.5 Pro’s reasoning into four target models. Qwen3.8 A95B’s answer overlap with the teacher jumped from 16.79% to 34.97%—an 18.18-point increase, especially large on STEM problems—while Kimi K3 moved only 4.54 points.
That suggests Qwen may have learned from GPT-5.5 Pro or a closely related model, but the experiment is behavioral evidence, not proof of a particular training pipeline. HN discussion raised shared-training-data and distillation alternatives, and debated whether the observed prefill effect says more about model imitation than internal reasoning.