Step-3
StepFunOpenApache 2.0
Step-3 is a vision-language model from StepFun. It ranks 106 of 167 on LMArena’s overall vision board with 1176±12 from 3,193 votes. Its best category is Homework (#101 of 118). License: Apache 2.0.
Key figures
- Developer
- StepFun
- License
- Apache 2.0
- First seen (our snapshots)
- 2026-10-26
No data: Captioning, Entity recognition, Creative writing, Humor, Chinese.
Who it suits
- Overall #106 of 167: lower half; consider it mainly for cost or deployment reasons.
- Strongest category: Homework (#101 of 118).
- Open weights (Apache 2.0): you can self-host it.
Neighbours on the overall board
| # | Model | Score | Votes | License | Δ vs this model |
|---|---|---|---|---|---|
| 103 |  GPT-4.1 mini (2025-04-14) OpenAI | 1181 | 40,201 | Proprietary | +5 |
| 104 |  Hunyuan Large Vision Tencent | 1180 | 1,274 | Proprietary | +4 |
| 105 | Hunyuan Vision 1.5 Thinking Tencent | 1180 | 2,424 | Proprietary | +4 |
| 106 | SStep-3 StepFun | 1176 | 3,193 | Apache 2.0 | |
| 107 |  Claude Opus 4 (20250514) Anthropic | 1176 | 2,172 | Proprietary | +0 |
| 108 | Claude Sonnet 4 (20250514) Anthropic | 1175 | 1,806 | Proprietary | -1 |
| 109 |  Mistral Medium 3.1 (2508) Mistral | 1172 | 41,611 | Proprietary | -4 |
Alternatives · More from StepFun
| # | Model | Score | Votes | License |
|---|---|---|---|---|
| 49 | SStep 5 Preview StepFun | 1265 | 1,649 | Proprietary |
| 102 | SStep-1o Turbo (202506) StepFun | 1186 | 1,895 | Proprietary |
| 127 | SStep-1o Vision 32K (High Res) StepFun | 1122 | 2,833 | Proprietary |
| 142 | SStep-1V 32K StepFun | 1046 | 1,534 | Proprietary |
Alternatives · Top open-weight models
| # | Model | Score | Votes | License |
|---|---|---|---|---|
| 24 |  GLM-5.3-Flash Z.ai | 1294 | 16,254 | MIT |
| 34 |  Kimi K2.6 Moonshot AI | 1282 | 16,368 | Modified MIT |
| 38 |  DeepSeek-V4.1-Flash-20260910 (Max) DeepSeek | 1278 | 4,100 | MIT |
| 39 |  Gemma 4 31B Google | 1276 | 40,219 | Apache 2.0 |
| 44 |  Qwen3.8-27B Alibaba | 1269 | 7,614 | Apache 2.0 |
How to read these scores
- If two scores differ by less than their ± intervals, the models are statistically close.
- Each category board is computed independently; do not compare scores across categories.
- Models with few votes move more; ranks of new entries may change.
FAQ
Where does Step-3 rank among vision models?
Overall rank 106 (1176); OCR rank 109 (1168); Diagrams rank 110 (1155); Homework rank 101 (1183); English rank 99 (1199) (LMArena, data as of Oct 26, 2026).
How good is Step-3 at OCR?
Step-3 ranks 109 of 123 on LMArena’s OCR board with 1168 (1,202 votes).
Is Step-3 open-weight?
LMArena lists its license as Apache 2.0, not proprietary; check the license text for commercial use.
How far is Step-3 from #1?
#1 overall is Claude Fable 5 (High) (1325); Step-3 has 1176, 149 points behind.
Sources
LMArena data published Oct 19, 2026.
Synced here Oct 26, 2026.
Price, context length and release date have no reliable public data here and are not shown; all figures come from snapshots, blanks are –.