实测：DeepSeek V4.1 Flash 在 DeepSeek Harness 上跑出 100% agent 基准

Kev @Kevrsub原推发布于 2026.10.24

在 X 上打开原推

I wanted to test what both did at default. On an agentic bench I use, DSH with Deepseek v4.1-Flash (jundots oMLX quant) was the first and only local model to date to score 100% on this bench. I didn't change anything with DSH or PI. Effort -50, reasoning budget of 7k. Same model with PI had a wider trajectory and lower capability ceiling. I was more focused on the bench results (quality) and DSH had a way longer wall clock time and higher tokens but I couldn't argue with the quality. Happy testing 😉