STEP3-VL-10B ranks first on MMBench.
StepFun Jan 2026 Open 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 | |
|---|---|---|---|---|
| WildBench | 86.0 | #1 of 4 | Model card | |
| MM-IFEval | 61.9% | #1 of 3 | Model card | |
| MultiChallenge | 62.6% | #14 of 36 | Model card | |
| HallusionBench | 64.9% | #9 of 23 | Model card | |
| MMLU-Pro | 76% | #97 of 151 | Model card | |
| Arena-Hard v2.0 | 58.6% | #17 of 24 | Model card | |
| IFEval | 82.2% | #91 of 126 | Model card | |
| Humanity's Last Exam | 10.7% | #131 of 147 | Model card | |
| IFBench | 43.3% | #65 of 70 | Model card | |
| SimpleVQA | 53.1% | #22 of 23 | Model card | |
| SuperGPQA | 50.4% | #25 of 26 | Model card |
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| AIME 2024 | 90.9% | #16 of 87 | Model card | |
| AIME 2025 | 87.7% | #63 of 143 | Model card | |
| GPQA Diamond | 70.8% | #168 of 284 | Model card | |
| Beyond AIME | 63.2% | #9 of 14 | Model card | |
| CNMO 2024 | 78.2% | #4 of 6 | Model card | |
| HMMT 2025 | 78.2% | #36 of 49 | Model card | |
| IMO-AnswerBench | 62.1% | #37 of 37 | Model card | |
| PhyX | 59.5% | #3 of 3 | Model card |
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| LiveCodeBench | 75.8% | #17 of 73 | Model card | |
| LiveCodeBench v6 | 75.8% | #39 of 76 | Model card | |
| Design2Code | 79.5 | #2 of 2 | Model card | |
| HumanEval-V | 66% | #1 of 1 | Model card |
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| ScreenSpot-v2 | 92.6% | #8 of 14 | Model card | |
| OSWorld-G | 59% | #3 of 5 | Model card | |
| ScreenSpot-Pro | 51.5% | #22 of 25 | Model card | |
| MMBench-GUI | 81.5% | #1 of 1 | Model card |
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| HealthBench | 44.7% | #29 of 30 | Model card |
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| MMBench | 92% | #1 of 9 | Model card | |
| MMMU | 78.1% | #23 of 97 | Model card | |
| ReMI | 67.3% | #1 of 4 | Model card | |
| MathVista | 84% | #18 of 64 | Model card | |
| MM-MT-Bench | 8.14 | #2 of 6 | Model card | |
| BLINK | 66.8% | #10 of 28 | Model card | |
| AI2D | 89.3% | #16 of 44 | Model card | |
| MMStar | 77.5% | #12 of 30 | Model card | |
| MathVerse | 75.7% | #5 of 12 | Model card | |
| OCRBench | 86.8% | #13 of 28 | Model card | |
| LogicVista | 66.9% | #6 of 12 | Model card | |
| All-Angles Bench | 57.2% | #3 of 6 | Model card | |
| We-Math | 73% | #6 of 11 | Model card | |
| MATH-Vision (MathVision) | 70.8% | #23 of 42 | Model card | |
| OmniSpatial | 51.6% | #3 of 5 | Model card | |
| DynaMath | 56.4% | #11 of 17 | Model card | |
| RealWorldQA | 74.4% | #20 of 30 | Model card | |
| MindCube | 62.8% | #2 of 3 | Model card | |
| MMMU-Pro | 64.1% | #50 of 72 | Model card | |
| CountBench | 88.8% | #15 of 20 | Model card | |
| MMSI-Bench | 32.2% | #4 of 5 | Model card | |
| CVBench | 83.5% | #9 of 11 | Model card | |
| CharXiv | 59.5% | #58 of 67 | Model card | |
| ERQA | 48.9% | #23 of 26 | Model card | |
| V* | 82.8% | #11 of 12 | Model card | |
| ZeroBench | 1.00% | #25 of 25 | Model card | |
| VisuLogic | 29.7% | #7 of 7 | Model card | |
| MMVP | 68.2% | #5 of 5 | Model card | |
| CountQA | 33.7% | #3 of 3 | Model card | |
| MIA-Bench | 92% | #2 of 2 | Model card | |
| PixMo-Count | 70.8% | #2 of 2 | Model card | |
| ViewSpatial-Bench | 46.1% | #2 of 2 | Model card | |
| CoreCognition | 66.7% | #1 of 1 | Model card | |
| CSVQA | 63.3% | #1 of 1 | Model card | |
| Do You See Me | 67.5% | #1 of 1 | Model card | |
| M3GIA | 78.4% | #1 of 1 | Model card | |
| OmniAI OCR Benchmark | 77% | #1 of 1 | Model card | |
| SpatialViz-Bench | 45.5% | #1 of 1 | Model card | |
| STARE | 61.8% | #1 of 1 | Model card |
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| LiveBench | 69.7% | #12 of 27 | Model card |
The benchmarks the lab chose to headline, with the scores it gave.
Every result in the longer documents the lab published with it. 52 of them aren't in the launch post.