Skip to content
Hbench

DeepSeek Harness

deepseek-harness

DeepSeek Harness's pass rate on every board it ran on. Scores on different boards don't add up, so this page doesn't add them.

On each board

SWE-bench Verified

swebench-verified@1.0
Claude Opus 5.50.2.0-rc.2, full, 500 tasks, 1 attempt, allowlist
Rank 1
2 harnesses
93.4%91.2–95.4DeepSeek Harness 0.2.0-rc.2 is the reference.

humaneval-plus@1.0 (unreviewed)

ghcr.io
GPT-6.1 Sol0.2.1-rc.2, full, 164 tasks, 1 attempt, open
Ranks 1–3
3 harnesses
96.3%93.3–98.8DeepSeek Harness 0.2.1-rc.2 is the reference.