Atlas

Benchmarks

← All benchmarks

MMLU-ProX

General QA · 2025-03-13

MMLU-ProX translates MMLU-Pro's ten-option questions into 29 typologically diverse languages, 11,829 questions per language, with a 658-question-per-language lite split. The headline figure averages accuracy across languages. Release reports often do not say whether the full or the lite split was scored, so no item count is declared.

Top models (higher is better)

ModelScore
Gemini 3 Pro Preview87.7
Qwen3.7-Max (source unspecified)87.0
Opus 4.686.1
Opus 4.585.7
Qwen3.7-Plus (source unspecified)85.4
Gemini 2.5 Pro84.7
Qwen3.5 397B A17B84.7
Qwen3.6 Plus (2026-04-02)84.7
GLM-5.183.9
DeepSeek-V4-Pro83.9
GPT-5.283.7
Kimi K2.683.7
o383.3
GLM-583.1
Kimi K2.582.3
Qwen3.5 122B A10B82.2
Qwen3.5-27B82.2
Qwen3 235B A22B Thinking 250781.0
Qwen3.5 35B A3B81.0
DeepSeek-R1-052880.6
Qwen3 VL 235B A22B Thinking80.6
Gemini 2.5 Flash80.2
Qwen3-235B-A22B80.0
Qwen3 235B A22B Instruct 250779.4
O4 Mini79.0
Qwen3 Max Thinking (2026-01-23)78.5
GPT-5 Mini78.5
Qwen3-VL-235B-A22B-Instruct77.8
Qwen3 VL 32B Thinking77.2
Qwen3 Next 80B A3B Instruct76.7
Qwen3.5-9B76.3
ChatGPT-4o Latest observed 2025-03-2776.2
Qwen3 VL 30B A3B Thinking76.1
DeepSeek-V3-032475.8
Qwen3 32B74.6
Kimi K2 Instruct74.5
gpt-oss-120b74.5
Qwen3 VL 32B Instruct73.4
Qwen3-30B-A3B73.1
Qwen3 30B A3B Instruct 250772.0
Loading Atlas data…