2026-11-09 · America/Los_Angeles · 社区动态 · #4

Same prompt, DeepSeek V4.1 Flash vs Qwen 3.8 FN: 94 tok/s vs 350 tok/s

I wanted to see how big of a difference there is between arguably two of the top MoE models that you can actually run locally, so I ran the same one shot prompt and left it unattended for a while. The prompt asks for a fully interactive V8 engine... yes, you're probably familiar with it and I'm aware this is not the best test and it only covers some of the capabilities of each model, and this particular test is likely on both of their training data, but still a fun experiment and the results were definitely interesting: DeepSeek V4.1 Flash 1017, UD IQ1_M (yup a "lobotomized" IQ1 quant, stop…

热度 46.4 / 100；排名与评分保留该期记录。