Step-3: benchmark results
StepFun's open-weight Step3 (July 2025), a 321-billion-parameter, 38-billion-active multimodal MoE reasoning model using Multi-Matrix Factorization Attention for efficient vision-language inference. Provider: StepFun. Released 2025-07-28. Access: Open.
Aggregate Rating 1628 ± 18, rank #365 of 1656 rated models, from 102 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| MathVision | 64.8 | Overall Accuracy (%) | 79.6 |
| FlagEval VQA - Geographic Reasoning | 58.2 | Score | 69.6 |
| Kagi LLM Benchmark | 62.3 | Accuracy (%) | 68.3 |
| FlagEval VQA - Text Understanding (Chinese) | 73 | Score | 65.2 |
| SuperCLUE-VLM (August 2025) - Overall | 63.64 | Score | 61.5 |
| FlagEval VQA - Visual Puzzles | 29.5 | Score | 60.9 |
| ReLE - Language and Instruction Following | 65.6 | Accuracy (%) | 56.8 |
| AGI-Eval Community - Learning | 85.6 | Accuracy (%) | 56.4 |
| AGI-Eval Community - Learning (English) | 90.53 | Accuracy (%) | 54.3 |
| AGI-Eval Community - Mathematical Reasoning | 74.96 | Accuracy (%) | 54.3 |
| Chatbot Arena (Text - Hard Prompts (English)) | 1404 | Arena Score | 54.1 |
| AGI-Eval Community - Learning (Chinese) | 78.52 | Accuracy (%) | 53.6 |
Interactive version: theaggregate.ai/model?slug=step-3 · How It Works · Data refreshed daily, snapshot 2026-10-20.