MMGist - STEM Reasoning: leaderboard

Metric: Accuracy (%) on the STEM Reasoning dimension (items kept from MathVista, MathVision, DynaMath and MME-SCI); MMGist: 7,262 curated items from 18 vision-language benchmarks after removing items answerable without the image, items nearly every model solves and items with faulty labels; eight samples per item at temperature 1.0, step-by-step answer in a boxed block, 16,384 output tokens, medium effort where configurable; higher is better. Source: arxiv.org. Saturation forecast: Around October 2027. 27 models tracked.

Top models

#ModelScore
1Qwen 3.6 Plus62.6
2Gemini 3.1 Pro (Preview) (Medium)61.2
3Seed 2.0 Pro57.6
4Qwen 3.6 35B A3B54.3
5Seed 2.0 Lite53.1
6Seed 2.0 Mini48.3
7Qwen 3.5 27B46.6
8GPT-5 (Medium)46.1
9Claude Sonnet 4.6 (Medium)44.6
10Gemini 3.1 Flash Lite (Medium)41.5
11Gemma 4 31B40.4
12Qwen 3.5 9B37.1
13Gemma 4 26B A4B36.8
14GPT-5 Mini (Medium)36.8
15Step3 VL 10B36.3

Interactive version: theaggregate.ai/benchmark?slug=mmgist-stem-reasoning · How It Works · Data refreshed daily, snapshot 2026-09-29.