Qwen 3.8 follows GPT-5.5 Pro reasoning prefills (gist.github.com)

The recent release of Qwen 3.8 marks a significant advancement in enhancing model performance through the use of reasoning prefills derived from GPT-5.5 Pro. The experiment compared the responses of various models—both prefilled with the initial reasoning of GPT-5.5 Pro and unprefilled—across a variety of problems, including STEM, non-STEM, and synthetic puzzles. Notably, Qwen 3.8 saw a remarkable increase in performance, with a 20.58 percentage point improvement in recall when utilizing the reasoning prefill, underscoring its potential to harness insights from advanced generative models. This development is particularly important for the AI and machine learning community, as it illustrates the efficacy of using prefills from superior models to boost the reasoning capabilities of subordinate models. The data demonstrates significant gains across all problem categories, especially in synthetic puzzles and STEM-related issues, indicating that Qwen may be effectively learning from the methodological approaches of GPT-5.5 Pro. The results not only highlight ongoing trends in model collaboration and performance enhancement but also explore a promising pathway for future model training methodologies in AI systems.