Benchmark Registry

DeepSeek AI

Benchmark results for 3 DeepSeek AI models from official publications. Updated 2026-10-07.

Established
Not available
Latest model
DeepSeek-V4-Flash (April 2026)

Benchmarks

107 results

ModelBenchmarkScoreSourceRegistry No.
DeepSeek-V4-Flash (Max)Riemann-bench October 2026 leaderboard10.4%surgehq.aiArchived copyRiemann-bench Pass@1Reported Benchmark leaderboardSurge AIIndependentRecord110001
DeepSeek-V4-Flash (max)AutomationBench 1.0.618.1%zapier.comArchived copyTask pass rateReported Benchmark leaderboardZapierIndependentZapier AutomationBench TeamRecord110001
DeepSeek-V4-Pro (max)GPQA Diamond — Epoch AI harness89.6%epoch.aiArchived copyGPQA Diamond AccuracyReported Evaluated Benchmark leaderboardEpoch AIIndependentRecord110002
DeepSeek-V4-Pro (max)Humanity's Last Exam Text only — Mercor single-shot34.4%mercor.comArchived copyHumanity's Last Exam AccuracyReported Benchmark leaderboardMercorIndependentRecord110002
DeepSeek-V4-Pro (none)GPQA Diamond — Epoch AI harness73.2%epoch.aiArchived copyGPQA Diamond AccuracyReported Evaluated Benchmark leaderboardEpoch AIIndependentRecord110002
DeepSeek-V4-Pro (max)DeepSWE 1.1 — Mercor mini-swe-agent, 500 steps9.1%mercor.comArchived copyDeepSWE Pass@1Reported Benchmark leaderboardMercorIndependentRecord110002
DeepSeek-V4-Pro (max)SWE Atlas Codebase QnA — Mercor mini-swe-agent, 250 steps25.0%mercor.comArchived copyMean pass@1Reported Benchmark leaderboardMercorIndependentRecord110002
DeepSeek-V4-Pro (max)GPQA Diamond — Mercor single-shot89.4%mercor.comArchived copyGPQA Diamond AccuracyReported Benchmark leaderboardMercorIndependentRecord110002
DeepSeek-V4-Pro (max)SimpleQA Verified Original — Epoch AI harness47.0%epoch.aiArchived copyAccuracyReported Evaluated Benchmark leaderboardEpoch AIIndependentRecord110002
DeepSeek-V4-Pro (max)Terminal-Bench 2.1 — Mercor Terminus 266.7%mercor.comArchived copyTerminal-Bench AccuracyReported Benchmark leaderboardMercorIndependentRecord110002
DeepSeek-V4-Pro (high)GPQA Diamond — Epoch AI harness90.9%epoch.aiArchived copyGPQA Diamond AccuracyReported Evaluated Benchmark leaderboardEpoch AIIndependentRecord110002
DeepSeek-V4-Pro (max)SWE-bench Verified — Mercor mini-swe-agent, 1000 steps75.4%mercor.comArchived copyResolved TasksReported Benchmark leaderboardMercorIndependentRecord110002
DeepSeek-V4-Pro (max)MMLU-Pro Original — Mercor single-shot86.5%mercor.comArchived copyAccuracyReported Benchmark leaderboardMercorIndependentRecord110002
DeepSeek-V4-Pro (Max)Riemann-bench October 2026 leaderboard38.4%surgehq.aiArchived copyRiemann-bench Pass@1Reported Benchmark leaderboardSurge AIIndependentRecord110002
DeepSeek-V4-Pro (max)BrowseComp 130-question subset — Mercor web research agent62.3%mercor.comArchived copyBrowseComp AccuracyReported Benchmark leaderboardMercorIndependentRecord110002
DeepSeek-V4-ProSWE Atlas Codebase QnA — Public; mini-swe-agent27.2%labs.scale.comArchived copyMean pass@1Reported Benchmark leaderboardScale AIIndependentRecord110002
DeepSeek-V4-Pro (max)SWE-bench Verified — Epoch AI harness77.6%epoch.aiArchived copyResolved TasksReported Evaluated Benchmark leaderboardEpoch AIIndependentRecord110002
DeepSeek-V3.2GPQA Diamond — Mercor single-shot85.2%mercor.comArchived copyGPQA Diamond AccuracyReported Benchmark leaderboardMercorIndependentRecord110003
DeepSeek-V3.2SWE Atlas Codebase QnA — Mercor mini-swe-agent, 250 steps5.9%mercor.comArchived copyMean pass@1Reported Benchmark leaderboardMercorIndependentRecord110003
DeepSeek-V3.2SWE-bench Verified — Mercor mini-swe-agent, 1000 steps66.0%mercor.comArchived copyResolved TasksReported Benchmark leaderboardMercorIndependentRecord110003
DeepSeek-V3.2Terminal-Bench 2.1 — Mercor Terminus 244.6%mercor.comArchived copyTerminal-Bench AccuracyReported Benchmark leaderboardMercorIndependentRecord110003
DeepSeek-V3.2DeepSWE 1.1 — Mercor mini-swe-agent, 500 steps0.6%mercor.comArchived copyDeepSWE Pass@1Reported Benchmark leaderboardMercorIndependentRecord110003
DeepSeek-V3.2MMLU-Pro Original — Mercor single-shot76.5%mercor.comArchived copyAccuracyReported Benchmark leaderboardMercorIndependentRecord110003
DeepSeek-V3.2BrowseComp 130-question subset — Mercor web research agent38.1%mercor.comArchived copyBrowseComp AccuracyReported Benchmark leaderboardMercorIndependentRecord110003
DeepSeek-V3.2 (High)Riemann-bench October 2026 leaderboard8.0%surgehq.aiArchived copyRiemann-bench Pass@1Reported Benchmark leaderboardSurge AIIndependentRecord110003
DeepSeek-V3.2Humanity's Last Exam Text only — Mercor single-shot22.3%mercor.comArchived copyHumanity's Last Exam AccuracyReported Benchmark leaderboardMercorIndependentRecord110003
DeepSeek-V3.2SWE-bench Pro Public — Scale AI SWE-Agent15.6%labs.scale.comArchived copyPublic tasks resolvedReported Benchmark leaderboardScale AIIndependentRecord110003
DeepSeek-V3.2SWE-bench Multilingual — Mercor mini-swe-agent, 1000 steps60.7%mercor.comArchived copyResolved TasksReported Benchmark leaderboardMercorIndependentRecord110003
DeepSeek-V3.2APEX-Agents Original21.3%mercor.comArchived copyPass rateReported Benchmark leaderboardMercorIndependentRecord110003
DeepSeek-V4-Pro (max)Terminal-Bench 4.0 — Vals AI mini-swe-agent11.1%vals.aiArchived copyTerminal-Bench AccuracyReported Benchmark leaderboardVals AIIndependentRecord110002
DeepSeek-V4-Pro (max)HLAB October 2026 evaluation3.8%vals.aiArchived copyHLAB Task Pass RateReported Benchmark leaderboardVals AIIndependentRecord110002
DeepSeek-V4-Pro (max)Vibe Code Bench 1.1 — OpenHands, October 2026 evaluation49.9%vals.aiArchived copyVibe Code Bench ScoreReported Benchmark leaderboardVals AIIndependentRecord110002
DeepSeek-V4-Pro (max)Finance Agent Benchmark 2 — October 2026 evaluation44.1%vals.aiArchived copyFinance Agent AccuracyReported Benchmark leaderboardVals AIIndependentRecord110002
DeepSeek-V4-Pro (max)Vals Index 2.138.6%vals.aiArchived copyVals Index GDP-weighted AccuracyReported Benchmark leaderboardVals AIIndependentRecord110002
DeepSeek-V3.2 (high)Vibe Code Bench 1.1 — OpenHands, October 2026 evaluation5.1%vals.aiArchived copyVibe Code Bench ScoreReported Benchmark leaderboardVals AIIndependentRecord110003
DeepSeek-V4-ProTerminal-Bench 2.1 — Vals AI Terminus 250.2%vals.aiArchived copyTerminal-Bench AccuracyReported Benchmark leaderboardVals AIIndependentRecord110002
DeepSeek-V4-Flash (max)DeepSWE 1.153.0%deepswe.datacurve.aiDeepSWE Pass@1Reported Benchmark leaderboardDatacurveIndependentRecord110001
DeepSeek-V4-Pro (max)DeepSWE 1.163.0%deepswe.datacurve.aiDeepSWE Pass@1Reported Benchmark leaderboardDatacurveIndependentRecord110002
DeepSeek-V4-Pro (max)MMLU-Pro Original — Vals AI harness87.2%vals.aiArchived copyAccuracyReported Benchmark leaderboardVals AIIndependentRecord110002
DeepSeek-V4-Pro (max)GPQA Diamond — Vals AI harness89.4%vals.aiArchived copyGPQA Diamond AccuracyReported Benchmark leaderboardVals AIIndependentRecord110002
DeepSeek-V4-Pro (max)SWE-bench Verified — Vals AI mini-swe-agent77.4%vals.aiArchived copyResolved TasksReported Benchmark leaderboardVals AIIndependentRecord110002
DeepSeek-V3.2 (high)GPQA Diamond — Vals AI harness80.3%vals.aiArchived copyGPQA Diamond AccuracyReported Benchmark leaderboardVals AIIndependentRecord110003
DeepSeek-V3.2 (none)GPQA Diamond — Vals AI harness76.3%vals.aiArchived copyGPQA Diamond AccuracyReported Benchmark leaderboardVals AIIndependentRecord110003
DeepSeek-V3.2 (high)SWE-bench Verified — Vals AI mini-swe-agent67.6%vals.aiArchived copyResolved TasksReported Benchmark leaderboardVals AIIndependentRecord110003
DeepSeek-V3.2 (thinking)MMLU-Pro Original — Vals AI harness84.9%vals.aiArchived copyAccuracyReported Benchmark leaderboardVals AIIndependentRecord110003
DeepSeek-V3.2MMLU-Pro Original — Vals AI harness83.1%vals.aiArchived copyAccuracyReported Benchmark leaderboardVals AIIndependentRecord110003
DeepSeek-V4-Flash (max)Toolathlon Verified50.9%toolathlon.xyzArchived copyTasks completedReported Benchmark leaderboardToolathlonIndependentToolathlon AuthorsRecord110001
DeepSeek-V4-Pro (max)Toolathlon Verified55.9%toolathlon.xyzArchived copyTasks completedReported Benchmark leaderboardToolathlonIndependentToolathlon AuthorsRecord110002
DeepSeek-V4-ProTerminal-Bench 2.0 — Vals AI Terminus 256.2%vals.aiArchived copyTerminal-Bench AccuracyReported Benchmark leaderboardVals AIIndependentRecord110002
DeepSeek-V3.2 (none)Terminal-Bench 2.0 — Vals AI Terminus 234.8%vals.aiArchived copyTerminal-Bench AccuracyReported Benchmark leaderboardVals AIIndependentRecord110003

DeepSeek AI models