According to a post on Hacker News, experiments regarding the inference generation patterns (reasoning prefill) of open models reported that Qwen shows a high correlation with the generation patterns of GPT-5.5 Pro.

In this experiment, the degree to which the teacher model's (GPT-5.5 Pro) responses were contained within the first 100 tokens of the target model's generated answers was measured. Forty-five questions, including STEM, non-STEM, and synthetic puzzles, were used for evaluation. As a result, Qwen showed an 18.18-point increase in similarity to GPT-5.5 Pro compared to its similarity to Opus 4.8 in previous experiments.

This data suggests the possibility that Qwen learned from GPT-5.5 Pro or related GPT models rather than Opus. Additionally, it was confirmed that Kimi K3 had the highest overlap rate with GPT-5.5 Pro, regardless of the presence of a prefill.

Source: