Distilling DeepSeek V4 Pro’s Thinking Style Into Qwen3.6-35B-A3B

Distilling DeepSeek V4 Pro’s Thinking Style Into Qwen3.6-35B-A3B 🌟Insights from Zhihu contributor & individual developer Lynn TL;DR: After reading ReAct, Lynn distilled the thinking style of DeepSeek V4 Pro into Qwen3.6-35B-A3B with LoRA. The result: GPQA-Diamond +7.6pp, unclosed empty answers down from 12 → 1, and average agent orchestration time down from 60s → 26s. The goal was not “more CoT,” but faster task decomposition, delegation, and verification. The starting point was simple: after reading Yao Shunyu’s ReAct work, the author wanted a local model that could act as an agent orchestrator. ReAct’s core idea is clear: let the model alternate between Thought and Action, instead of going to either extreme