Step3 VL 10B: benchmark results
StepFun's Apache-2.0 10B vision-language model pairing a perception encoder with a Qwen3-8B decoder, rivaling multimodal models 10-20x its size (January 2026). Provider: StepFun. Released 2026-01-20. Access: Open.
Aggregate Rating 1587 ± 11, rank #485 of 1656 rated models, from 177 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| LLM Stats (MMBench) | 91.8 | Score (%) | 100 |
| EXPLORE-Bench - Long Sequences | 47.09 | Unified score (0-100): 0.25 x object coverage + 0.35 x attri | 88.9 |
| EXPLORE-Bench - Object Coverage | 60.61 | Object-level coverage (0-100): soft Sentence-BERT matching o | 88.9 |
| MathVision | 75.95 | Overall Accuracy (%) | 87.9 |
| LLM Stats (MathVista) | 84 | Score (%) | 86.8 |
| DiffCap-Bench - Camera and Viewpoint | 77.5 | Macro-averaged F1* (%) on camera movement and viewpoint chan | 86.7 |
| EXPLORE-Bench | 48.87 | Unified score (0-100): 0.25 x object coverage + 0.35 x attri | 85.2 |
| EXPLORE-Bench - Medium Sequences | 48.01 | Unified score (0-100): 0.25 x object coverage + 0.35 x attri | 85.2 |
| EXPLORE-Bench - Relation Score | 2.74 | Relation-level score (0-5): a Qwen3-8B scorer rates the pred | 85.2 |
| EXPLORE-Bench - Short Sequences | 49.87 | Unified score (0-100): 0.25 x object coverage + 0.35 x attri | 85.2 |
| SIS-Bench - Spatio-Temporal Consistency | 54.4 | Accuracy (%; 241 spatio-temporal consistency questions; four | 80 |
| LLM Stats (Multi-Challenge) | 62.6 | Score (%) | 74.2 |
Interactive version: theaggregate.ai/model?slug=step3-vl-10b · How It Works · Data refreshed daily, snapshot 2026-10-20.