Open LMM Reasoning - WeMath - Three-step(S3): leaderboard
The "WeMath - Three-step(S3)" subtask of Open LMM Reasoning. A task from the Open LMM Reasoning leaderboard, which evaluates large multimodal models on multi-step visual, mathematical and logical reasoning problems.
Metric: Accuracy (%). Source: huggingface.co. Saturation forecast: Estimated already saturated. 83 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Doubao-1.5-Pro | 70.9 | #287 |
| 2 | GPT-4.1 (2025-04-14) | 61.8 | #250 |
| 3 | Gemini 2.0 Flash | 60.6 | #342 |
| 4 | GPT-4o ChatGPT | 58.8 | |
| 5 | Claude 3.7 Sonnet | 56.4 | #249 |
| 6 | Taichu-VLR-7B | 56.4 | #526 |
| 7 | Claude 3.5 Sonnet (20241022) | 54.5 | #291 |
| 8 | Gemini 1.5 Pro (002) | 53.9 | #349 |
| 9 | QVQ-72B-Preview | 53.9 | #440 |
| 10 | InternVL2.5-8B-BoN-8 | 53.9 | #639 |
| 11 | InternVL3-38B | 52.7 | #405 |
| 12 | InternVL3-78B | 50.3 | #352 |
| 13 | InternVL3-14B | 49.7 | #505 |
| 14 | GPT-4.1 Nano | 49.1 | #731 |
| 15 | VLAA-Thinker-Qwen2.5VL-7B | 49.1 | #555 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=open-lmm-reasoning-wemath-three-step-s3 · How It Works · Data refreshed daily, snapshot 2026-10-20.