Atlas

Benchmarks

← All benchmarks

CMATH

Math · 2023-06-29

CMATH is a Chinese elementary-school math word-problem benchmark of 1,698 problems from grades one to six, graded by final-answer match. Scores report accuracy; shot count is recorded per row. Higher is better.

Top models (higher is better)

ModelScore
DeepSeek-V4-Flash-Base93.6
DeepSeek-V4-Pro-Base90.9
Kimi K2 Base90.3
DeepSeekMath 7B RL88.8
Llama 4 Maverick88.1
Kimi Linear 48B A3B Base85.5
DeepSeekMath 7B Instruct84.6
Qwen1.5 72B Chat82.8
DeepSeek-V2-Chat81.9
DeepSeek LLM 67B Chat80.3
Llama 3 70B Instruct79.2
DeepSeek-V278.7
Qwen1.5 72B77.8
Llama 3.1 405B77.3
Mixtral 8x22B Instruct75.1
Llama 3 70B73.9
Mixtral 8x22B72.3
DeepSeekMath 7B Base71.7
DeepSeek-V2-Lite-Chat71.7
DeepSeekMoE 16B Chat67.9
DeepSeek LLM 67B Base63.0
DeepSeek-V2-Lite58.4
Mistral 7B v0.1 Base44.9
DeepSeekMoE 16B Base40.4
DeepSeek LLM 7B Base34.5
Loading Atlas data…