MathVista
14 models tested · Updated 2026-04-14 · Verified sources only
EXAONE 4.5 33B leads at 85.0%
1
LG AI Research · HuggingFace/LGAI-EXAONE · 2026-04-14
Beats GPT-5 mini (79.1). Strong visual math reasoning for 33B open-weight VLM.
85.0%
2
arxiv · arxiv/2604.08539 · 2026-04-09
8B open-weight multimodal model trained with GRPO+GDPO. Competitive with Gemini 2.5 Pro on DocVQA and chart understanding. New SOTA for open-weight VLMs on MMMU.
82.7%
3
arxiv · arxiv/2603.03975 · 2026-03-04
Compact 15B open-weight multimodal reasoning model from Microsoft. Achieves competitive VLM performance with much less compute via careful data curation and dynamic-resolution encoders.
81.8%
4
UCLA · arxiv/2604.08539 · 2026-04-09
Open-weight 8B VLM surpassing GPT-4o (63.8%) by 15.7 points on MathVista. Demonstrates strong visual math reasoning.
79.5%
5
arxiv · arxiv/2604.08539 · 2026-04-09
Introduces Gaussian GRPO (G2RPO), replacing standard linear scaling in GRPO with non-linear distributional matching. OpenVLThinkerV2-7B achieves new SOTA for open-source 7B models on MMMU (71.6%), Mat
79.5%
6
Vero Team · arxiv/2604.04917 · 2026-04-06
Slightly below thinking base (81.4%) but more consistent across all 30 benchmarks.
79.2%
7
Vero Team · arxiv/2604.04917 · 2026-04-06
+1.5 over base. Consistent improvement across visual reasoning tasks via task-routed rewards.
78.7%
8
arxiv · arxiv/2604.03128 · 2026-04-06
Proposes RLSD combining self-distillation magnitude with RLVR direction. On Qwen3-VL-8B, achieves best avg accuracy across 5 multimodal reasoning benchmarks, outperforming GRPO by 2.32% on average.
78.1%
9
arxiv · arxiv/2603.03975 · 2026-03-04
Compact 15B open-weight multimodal reasoning model from Microsoft. Achieves competitive VLM performance with much less compute via careful data curation and dynamic-resolution encoders.
77.1%
10
arxiv · arxiv/2603.03975 · 2026-03-04
Compact 15B open-weight multimodal reasoning model from Microsoft. Achieves competitive VLM performance with much less compute via careful data curation and dynamic-resolution encoders.
75.2%
11
arxiv · arxiv/2604.08539 · 2026-04-09
8B open-weight multimodal model trained with GRPO+GDPO. Competitive with Gemini 2.5 Pro on DocVQA and chart understanding. New SOTA for open-weight VLMs on MMMU.
74.2%
12
arxiv · arxiv/2603.03975 · 2026-03-04
Compact 15B open-weight multimodal reasoning model from Microsoft. Achieves competitive VLM performance with much less compute via careful data curation and dynamic-resolution encoders.
67.1%
13
arxiv · arxiv/2604.08539 · 2026-04-09
8B open-weight multimodal model trained with GRPO+GDPO. Competitive with Gemini 2.5 Pro on DocVQA and chart understanding. New SOTA for open-weight VLMs on MMMU.
63.8%
14
arxiv · arxiv/2603.03975 · 2026-03-04
Compact 15B open-weight multimodal reasoning model from Microsoft. Achieves competitive VLM performance with much less compute via careful data curation and dynamic-resolution encoders.
57.4%