Testing qwen3.8 27b, beating deepseek-v4-flash without deep thinking mode turned on, please check the description
28.5K
8
2026-08-15 22:32:15
Personal opinion, for reference only
115
16
98
26
课程总结
收起全部章节
关闭课程总结
暂无课程总结
Tested with deepseek harness, qwen3.8-28b-Q6K, turned on deep thinking to make a Xiangqi web game, 200k context directly keeps thinking, the context is insufficient.
Conclusion 1: When deep thinking is on without limiting thinking length, it gets crushed by deepseek-v4-flash, because the model keeps thinking, and 200k context is not enough. My two 3080s can only hold 200k of context without quantized KV cache.
Conclusion 2: When deep thinking is off, it crushes deepseek-v4-flash, because what deepseek-v4-flash produces is garbage, while what qwen3.8-28b-Q6K produces is basically normal.
