DeepSeek Harness 实测 V4.1 Flash-1017：9k 推理预算下 16/16 满分，6 次中 3

Kev @Kevrsub原推发布于 2026.10.23

在 X 上打开原推

Final testing of Deepseek v4.1-Flash-1017 in the Deepseek harness. Reasoning 40 with a reasoning budget of 9k. This Bench hit 16/16 (100%) for 1 run of 6. This ceiling is real and it is higher but the output is less consistent with the range widening. Run times are longer and the full runtime was needed on an M3 Ultra. Timeouts were set to 210 minutes and this was hit 3 out of 6 times. This is the best scoring model I have benched locally to date. Again - a reasoning effort using PI is not equivalent to the same reasoning effort in DSH.