DeepSeek V4 Flash 0731 ranks first on Long-Horizon-Terminal-Bench (LHTB).
DeepSeek Jul 2026 Open weights Mid-size
Its best score on each benchmark, and where that ranks among every model with a score there. We rank each model once, at its best setting.
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| BullshitBench | 40% | #37 of 107 | Official leaderboard | |
| Longform Creative Writing (EQ-Bench) | 61.2 | #60 of 119 | Official leaderboard | |
| OpenAI MRCR v2 | 44.5% | #24 of 47 | Official leaderboard | |
| BullshitBench V2 | 39% | #66 of 129 | Official leaderboard | |
| Creative Writing v3 | 67.5 | #62 of 120 | Official leaderboard | |
| SimpleQA Verified | 33.6% | #66 of 94 | Epoch AI run | |
| Judgemark v4 | 30.8 | #35 of 40 | Official leaderboard | |
| DesignArena ASCII Art | 1,093 | #69 of 77 | Official leaderboard |
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| GPQA Diamond | 91% | #36 of 229 | Epoch AI run | |
| DTBench | 90.9% | #43 of 180 | Official leaderboard | |
| OTIS Mock AIME 2024-2025 | 94.4% | #53 of 212 | Epoch AI run | |
| ProofBench | 39% | #19 of 59 | Official leaderboard | |
| ARC-AGI-2 | 61.4% | #32 of 98 | Official leaderboard | |
| ARC-AGI | 89% | #37 of 95 | Official leaderboard | |
| FrontierMath Tiers 1-3 v2 | 57.5% | #44 of 86 | Official leaderboard | |
| ProofBench v1.1 | 56% | #28 of 52 | Official leaderboard | |
| SimpleBench | 61.1% | #10 of 18 | Collected by Epoch AI, source not named | |
| FrontierMath Tier 4 v2 | 24.4% | #44 of 68 | Official leaderboard | |
| FrontierChallenge | 12.4% | #8 of 11 | Official leaderboard | |
| BioMysteryBench | 64.4% | #24 of 30 | Official leaderboard |
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| WebDev Arena | 1,580 | #25 of 105 | arena.ai, via Epoch AI | |
| WeirdML | 63% | #27 of 112 | Official leaderboard | |
| ALE-Bench | 1,686 | #46 of 126 | Official leaderboard | |
| Vibe Code Bench | 74.7% | #44 of 120 | Official leaderboard | |
| NL2Repo-Bench | 54.2% | #12 of 30 | Lab model card | |
| Code Migration | 38.6% | #33 of 78 | Official leaderboard | |
| DesignArena Website | 1,231 | #66 of 151 | Official leaderboard | |
| DesignArena UI Component | 1,228 | #63 of 139 | Official leaderboard | |
| DesignArena Game Dev | 1,199 | #66 of 143 | Official leaderboard | |
| DesignArena 3D Design | 1,190 | #64 of 136 | Official leaderboard | |
| DesignArena Code | 1,220 | #69 of 143 | Official leaderboard | |
| DesignArena SVG | 1,171 | #52 of 103 | Official leaderboard | |
| GameCraft-Bench | 40.6% | #8 of 15 | Official leaderboard | |
| DesignArena Data Visualization | 1,170 | #78 of 118 | Official leaderboard | |
| DSBench (DeepSeek internal) | 68.7% | #2 of 3 | Lab model card | |
| DeepSWE | 54.4% | #46 of 56 | Lab model card | |
| LoopArena | 28.3% | #5 of 6 | Official leaderboard | |
| IOI v2 (Vals) | 32.7% | #40 of 45 | Official leaderboard | |
| ReactBench | 21.5% | #26 of 28 | Official leaderboard |
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| Long-Horizon-Terminal-Bench (LHTB) | 60.2% | #1 of 25 | Official leaderboard | |
| Mystery Game Puzzles | 34% | #28 of 84 | Epoch AI run | |
| GAIA | 66.5% | #22 of 64 | Official leaderboard | |
| Toolathlon-Verified | 70.7% | #9 of 25 | Official leaderboard | |
| Terminal-Bench 2.1 | 82.7% | #24 of 56 | Lab model card | |
| Chess Puzzles | 32% | #52 of 93 | Epoch AI run | |
| BabaIsBench | 75% | #20 of 32 | Official leaderboard | |
| RuneBench | 4.20 | #51 of 76 | Official leaderboard | |
| AutomationBench | 25.1% | #39 of 47 | Lab model card | |
| DiG-bench | 11.4% | #6 of 7 | Official leaderboard |
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| Harvey LAB (Legal Agent Benchmark) | 8.33% | #26 of 80 | Official leaderboard | |
| CorpFin v2 | 61.9% | #53 of 119 | Official leaderboard | |
| Vals Index v2 | 53.6% | #30 of 65 | Official leaderboard | |
| Tax Agent Bench | 28.1% | #39 of 84 | Official leaderboard | |
| FoodTruck Bench | $24,039 | #29 of 59 | Official leaderboard | |
| PolicyBench | 82.9% | #26 of 50 | Official leaderboard | |
| MedCode | 41.4% | #51 of 96 | Official leaderboard | |
| MedScribe | 80.4% | #54 of 97 | Official leaderboard | |
| Excel Modeling Benchmark | 57% | #43 of 76 | Official leaderboard | |
| Vals Finance Agent v2 | 49.5% | #47 of 80 | Official leaderboard | |
| Legal Research Bench | 30.3% | #56 of 93 | Official leaderboard | |
| TRLawBench | 8.93 | #21 of 34 | Official leaderboard | |
| TaxEval v2 | 70.7% | #82 of 129 | Official leaderboard | |
| LegalBench | 77.7% | #100 of 138 | Official leaderboard | |
| Vals Index v2.1 | 48% | #36 of 49 | Official leaderboard | |
| Agents' Last Exam | 25.2% | #22 of 26 | Lab model card |
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| MineBench | 1,482 | #40 of 75 | Official leaderboard | |
| Code2Video Benchmark | 1,472 | #13 of 19 | Official leaderboard | |
| VoxelBench | 1,444 | #41 of 56 | Official leaderboard |
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| CyberBench (Vals) | 77.5% | #6 of 26 | Official leaderboard | |
| CyberBench v1.1 (Vals) | 72.1% | #17 of 44 | Official leaderboard | |
| nocot-bench (NCRI) | 75.5 | #73 of 185 | Official leaderboard | |
| PatchEval | 80.4% | #3 of 7 | Official leaderboard | |
| PatchEval-Verified | 80.4% | #3 of 7 | Official leaderboard | |
| CyberGym | 76.7% | #21 of 34 | Lab model card | |
| SecIT Bench | 69% | #19 of 26 | Official leaderboard | |
| CyberGym-E2E | 28.4% | #5 of 5 | Official leaderboard |
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| LMCA | 41.7% | #51 of 147 | Official leaderboard | |
| Conceptual Reasoning Index | 54.4 | #55 of 151 | Official leaderboard | |
| ACCoRD | 61.7 | #86 of 152 | Official leaderboard |
The Self-Reported Index fits one score to the 6 capability results DeepSeek reported for this model, on the scale of the Epoch Capabilities Index. Labs choose which benchmarks to show and how to run them, so read it as the lab's claim, not an independent measurement. The 90% interval runs from 151.4 to 158.7. How we compute it.
Where the benchmark's own leaderboard, or Epoch AI's run, also has this model. Gaps mostly come from different harnesses, effort levels and task subsets. Every pair.
| Benchmark | Lab says | Board says | Gap |
|---|---|---|---|
| Toolathlon-Verified | 70.3% Toolathlon-Verified | 70.7% Leaderboard, max | −0.4 pts |
DeepSeek's models in color, against every other model on the Epoch Capabilities Index, by release date.
Source: Epoch Capabilities Index, Epoch AI (CC BY 4.0)
The benchmarks the lab chose to headline, with the scores it gave.