Gemini 3.1 Flash-Lite ranks first on StructEval.
Google DeepMind Mar 2026 Closed weights Small
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 | |
|---|---|---|---|---|
| MMLU-Pro | 86.2% | #23 of 212 | Official leaderboard | |
| SEA-HELM | 75.3 | #13 of 62 | Official leaderboard | |
| Wer wird Millionär? LLM Benchmark | 114156.0 | #16 of 76 | Official leaderboard | |
| MMMLU | 88.9% | #15 of 62 | Model card | |
| LMArena (Elo) | 1,433 | #88 of 333 | Official leaderboard | |
| Judgemark v2.1 | 80.9 | #15 of 42 | Official leaderboard | |
| Vectara Hallucination Leaderboard (2025 dataset) | 8.20% | #41 of 104 | Official leaderboard | |
| SimpleQA Verified | 43.3% | #9 of 22 | Model card | |
| Phare | 67.5% | #27 of 62 | Official leaderboard | |
| MultiChallenge | 60.6% | #16 of 35 | Official leaderboard | |
| FACTS Parametric | 48.4% | #21 of 45 | Official leaderboard | |
| Judgemark v4 | 58.8 | #20 of 40 | Official leaderboard | |
| FACTS Multimodal | 39.4% | #22 of 39 | Official leaderboard | |
| FACTS Search | 66.8% | #18 of 31 | Official leaderboard | |
| MultiNRC | 25% | #23 of 38 | Official leaderboard | |
| Humanity's Last Exam | 8.64% | #30 of 47 | Official leaderboard | |
| MRCR v2 | 60.1% | #18 of 27 | Model card | |
| Longform Creative Writing (EQ-Bench) | 49.0 | #83 of 119 | Official leaderboard | |
| FACTS Benchmark Suite | 39.4% | #29 of 40 | Official leaderboard | |
| BullshitBench V2 | 11% | #105 of 122 | Official leaderboard | |
| BullshitBench | 9.09% | #99 of 107 | Official leaderboard | |
| Extended NYT Connections | 8.20% | #93 of 97 | Official leaderboard |
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| GPQA Diamond | 81.8% | #86 of 219 | Epoch AI | |
| OTIS Mock AIME 2024-2025 | 80% | #89 of 202 | Epoch AI | |
| ForecastBench | 62.5 | #42 of 86 | Official leaderboard | |
| DTBench | 61.3% | #99 of 178 | Official leaderboard | |
| CritPt | 1.14% | #93 of 160 | Epoch AI | |
| FrontierMath Tiers 1-3 v2 | 27.7% | #70 of 89 | Epoch AI | |
| BixBench3 | 0.00 | #20 of 20 | Official leaderboard |
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| StructEval | 81.2% | #1 of 28 | Official leaderboard | |
| LiveCodeBench | 72% | #24 of 73 | Model card | |
| WeirdML | 52.2% | #46 of 112 | Official leaderboard | |
| ALE-Bench | 798 | #70 of 121 | Epoch AI | |
| Code Arena | 1,256 | #110 of 124 | Official leaderboard | |
| WebDev Arena | 1,254 | #94 of 105 | Epoch AI | |
| Code Migration | 4.61% | #73 of 78 | Official leaderboard | |
| Image-to-WebDev Arena | 1,336 | #54 of 55 | Official leaderboard | |
| Vibe Code Bench | 0% | #113 of 115 | Official leaderboard | |
| FrontierCode | 4.84% | #12 of 12 | Official leaderboard |
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| SnakeBench | 31 | #34 of 183 | Official leaderboard | |
| PinchBench 2.0 | 80.5% | #16 of 57 | Official leaderboard | |
| PACT | 1,509 | #16 of 26 | Official leaderboard | |
| LLM Chess | -133 | #82 of 124 | Official leaderboard | |
| DeepResearch Bench | 37.3% | #21 of 24 | Epoch AI | |
| MCP Atlas | 57.1% | #30 of 34 | Official leaderboard | |
| UniClawBench | 15.5% | #9 of 10 | Official leaderboard | |
| Toolathlon | 14.2% | #47 of 50 | Official leaderboard | |
| SkillsBench | 20.1% | #21 of 21 | Official leaderboard |
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| MortgageTax | 68% | #20 of 86 | Official leaderboard | |
| MedCode | 47.6% | #31 of 96 | Official leaderboard | |
| LegalBench | 83.8% | #49 of 138 | Official leaderboard | |
| TaxEval v2 | 71.8% | #70 of 129 | Official leaderboard | |
| CorpFin v2 | 59.4% | #75 of 119 | Official leaderboard | |
| Vals Finance Agent | 46.1% | #33 of 49 | Official leaderboard | |
| CaseLaw v2 | 55% | #36 of 53 | Official leaderboard | |
| PolicyBench | 79.3% | #38 of 50 | Official leaderboard | |
| CADGenBench | 24.9% | #9 of 11 | Official leaderboard | |
| Harvey LAB (Legal Agent Benchmark) | 0% | #68 of 80 | Official leaderboard | |
| Vals Finance Agent v2 | 30% | #73 of 80 | Official leaderboard | |
| MedScribe | 63.9% | #91 of 98 | Official leaderboard | |
| Legal Research Bench | 3.37% | #84 of 88 | Official leaderboard | |
| Tax Agent Bench | 3.11% | #77 of 79 | Official leaderboard | |
| Excel Modeling Benchmark | 8.63% | #76 of 76 | Official leaderboard |
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| GeoBench | 83% | #11 of 48 | Official leaderboard | |
| Video-MMMU | 84.8% | #8 of 30 | Model card | |
| VISTA | 46.9% | #20 of 54 | Official leaderboard | |
| MMMU-Pro | 76.8% | #27 of 72 | Model card | |
| Vision Arena | 1,235 | #58 of 136 | Official leaderboard | |
| 1H-VideoQA | 78.2% | #9 of 20 | Official leaderboard | |
| ZeroBench | 5.00% | #44 of 66 | Official leaderboard | |
| ParseBench | 58.3% | #28 of 42 | Official leaderboard | |
| CharXiv | 73.2% | #45 of 67 | Model card | |
| EnigmaEval | 3.04% | #29 of 41 | Epoch AI |
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| BrowseSafe++ | 80.9% | #8 of 15 | Official leaderboard | |
| FORTRESS | 51% | #52 of 66 | Official leaderboard | |
| PCT (Polarized Contrastive Pairs) | 51.9% | #37 of 46 | Official leaderboard |
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| Political Science LLM Benchmark | 68.1% | #6 of 29 | Official leaderboard | |
| SAGE (Vals) | 49.5% | #20 of 81 | Official leaderboard | |
| LMCA | 35% | #70 of 139 | Epoch AI | |
| ACCoRD | 64.1 | #76 of 150 | Official leaderboard | |
| Conceptual Reasoning Index | 46.1 | #81 of 149 | Official leaderboard | |
| CAIS Vision Capabilities Index | 44.7 | #39 of 53 | Official leaderboard | |
| CAIS Risk Index | 61.0 | #51 of 65 | Official leaderboard | |
| CAIS Text Capabilities Index | 18.1 | #53 of 66 | Official leaderboard | |
| Vals Multimodal Index | 41.4% | #31 of 33 | Official leaderboard |
Google DeepMind's models in color, against every other model on the Epoch Capabilities Index, by release date.
The benchmarks the lab chose to headline, with the scores it gave.
Every result in the longer documents the lab published with it. 5 of them aren't in the launch post.