MATH-Vision Leaderboard 2026 — Results Across 9 Real AI Models

MATH-Vision leaderboard

MATH-Vision

9 models tested · Updated 2026-04-02 · Verified sources only

      Gemma 4 31B leads at 85.6%
    

Multimodal math reasoning from images.

85.6%

MoE 25.2B total, 3.8B active.

82.4%

StepFun · arxiv/2601.09668 · 2026-01-14

10B model achieves 75.95% on MATH-Vision, competitive with Gemma 4 31B (85.6) at 1/3 the params.

75.95%

Vero Team · arxiv/2604.04917 · 2026-04-06

Highest MATH-Vision among all Vero variants. Thinking backbone excels on visual math.

63.5%

4.5B effective params.

59.5%

Vero Team · arxiv/2604.04917 · 2026-04-06

+2.6 over MiMoVL base. Surpasses MiMoVL-7B-RL which uses proprietary recipe.

59.4%

Vero Team · arxiv/2604.04917 · 2026-04-06

+5.1 over base. Open RL matches proprietary VLM training pipelines on visual math.

59.0%

Research · arxiv/2604.08539 · 2026-04-09

8B model trained with G2RPO reaches new open-source SOTA on MATH-Vision, beating larger models.

53.4%

2.3B effective params.

52.4%