RewardBench 2 Math — leaderboard
Metric: Accuracy (%). Source: huggingface.co. 52 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Qwen 3.6 Plus | 91.8 |
| 2 | Qwen 3.5 397B A17B | 91.8 |
| 3 | Kimi K2.6 | 91.26 |
| 4 | Qwen 3.5 122B A10B | 90.71 |
| 5 | NVIDIA-Nemotron-3-Super-120B-A12B-FP8 | 90.57 |
| 6 | MiMo-V2-Flash | 90.44 |
| 7 | GLM-4.6 FP8 | 90.44 |
| 8 | Qwen 3.6 35B A3B | 90.16 |
| 9 | Gemini 3.1 Pro (Preview) | 89.89 |
| 10 | Kimi K2 (Thinking) | 89.62 |
| 11 | GLM-5 FP8 | 89.62 |
| 12 | Qwen 3 Next 80B A3B (Thinking) | 89.34 |
| 13 | Step 3.5 Flash | 89.34 |
| 14 | GPT-5.5 | 89.21 |
| 15 | GPT-OSS-120B | 89.21 |
Interactive version: theaggregate.ai/benchmark?slug=rewardbench-2-math · How the rankings work · Data refreshed daily, snapshot 2026-07-22.