GSM8K-V: leaderboard

Each model solves all 1,319 GSM8K test problems redrawn as two to eleven images, seeing only the images and a format prompt, and the score is the share of correct final numbers. Unlike GSM8K and MathVista, no problem text is given, so the quantities and steps must be read from the scenes.

Metric: Accuracy (%). Source: arxiv.org. Saturation forecast: Around February 2027. 21 models tracked. Updated 17 Nov 2026.

Top models

#ModelSettingScoreOverall rank
1Gemini 3 Pro (Preview)59.06#120
2Kimi K2.647.34#191
3GLM-4.6V37.78#710
4Llama 4 Maverick Instruct31.69#1151
5GPT-5Minimal30.1#177
6Qwen 3 VL 30B A3B (Thinking)28.58#728
7GLM-4.5V28.28#758
8Qwen 3 VL 30B A3B Instruct27.9#894
9InternVL3.5-241B-A28B27.9#720
10Ovis2.5-9B27.45#1237
11Llama 4 Scout Instruct26.69#1575
12Qwen 3 VL 8B (Thinking)26.46#914
13Qwen 3 VL 8B Instruct26#1025
14Step-325.85#739
15InternVL3.5-30B-A3B22.82#1265

No result here: #1 Claude Fable 5.2, #2 GPT-6.2 Astra, #3 Gemini 4 Carbon.

Interactive version: theaggregate.ai/benchmark?slug=gsm8k-v · How It Works · Data refreshed daily, snapshot 2026-11-17.