// BENCHMARK

MATH benchmark

AI model leaderboard for the MATH benchmark. Compare how large language models score on MATH, see the full ranking, and understand what this AI benchmark measures. DeepSeek-V3 currently leads with 89.7%. 12,500 high-school competition math problems with step-by-step solutions.

Leaderboard

# Model Organization Score Variant Source
#1 DeepSeek-V3 DeepSeek 89.7% EM official ↗
#2 dots.llm1 Rednote 85% EM official ↗
#3 DeepSeek-V3 DeepSeek 84.6% cited-minimax-text official ↗
#4 Gemini 1.5 Pro Google DeepMind 84.6% 002-cited-minimax-text official ↗
#5 Gemini 2.0 Flash Google DeepMind 83.9% exp-cited-minimax-text official ↗
#6 Qwen2.5 Instruct (72B) Qwen 81.8% instruct-cited-minimax-text official ↗
#7 MiniMax-Text-01 MiniMax 77.4% from-hf-readme official ↗
#8 GPT-4o (Nov 2024) OpenAI 76.6% cited-minimax-text official ↗
#9 GPT-4 (Jun 2023) OpenAI 76.6% cited-nvlm official ↗
#10 Qwen2.5 Instruct (72B) Qwen 74.3% instruct-cited-tulu3 official ↗
#11 Claude 3.5 Sonnet Anthropic 74.1% 1022-cited-minimax-text official ↗
#12 Llama 3.1-405B Meta AI 73.8% instruct-cited-minimax-text official ↗
#13 Gemma 4 31B Google DeepMind 73.8% official ↗
#14 Llama 3.1-405B Meta AI 73.8% instruct-cited-nvlm official ↗
#15 NVLM-D 72B NVIDIA 73.1% megatron-from-hf-readme official ↗
#16 DeepSeek-V3 DeepSeek 72.5% cited-tulu3 official ↗
#17 Claude 3.5 Sonnet Anthropic 71.1% cited-nvlm official ↗
#18 GPT-4o mini OpenAI 70.2% cited-phi35moe official ↗
#19 GPT-4o mini OpenAI 70.2% cited-aria official ↗
#20 Mistral Small 3.2 Mistral AI 69.42% instruct-from-hf-readme official ↗
#21 Mistral Small 3.1 Mistral AI 69.3% instruct-cited-msmall32 official ↗
#22 GPT-4o (Nov 2024) OpenAI 68.8% cited-tulu3 official ↗
#23 Llama 3.1-70B Meta AI 68% instruct-cited-nvlm official ↗
#24 Gemini 1.5 Pro Google DeepMind 67.7% aug-2024-cited-nvlm official ↗
#25 Tulu 3 405B Allen Institute for AI 67.3% instruct-4-shot-flex official ↗
#26 Llama 3.1-405B Meta AI 66.6% instruct-cited-tulu3 official ↗
#27 Phi-4 Mini Microsoft 64% instruct-0-shot-cot official ↗
#28 Qwen2.5-3B Qwen 61.7% cited-phi4-mini official ↗
#29 Qwen2.5-7B Qwen 60.4% cited-phi4-mini official ↗
#30 Qwen-72B Qwen 59.7% instruct-cited-nvlm official ↗
#31 Phi-3.5-MoE Microsoft 59.5% instruct-0-shot-cot official ↗
#32 Nemotron-H 56B NVIDIA 59.42% base-4-shot-cot official ↗
#33 Hermes 3 405B Nous Research 58.4% cited-tulu3 official ↗
#34 Nemotron-H 47B NVIDIA 57.4% base-4-shot-cot official ↗
#35 Llama 3.1-70B Meta AI 56.4% instruct-cited-tulu3 official ↗
#36 Llama 3.2 11B Meta AI 51.9% cited-aria official ↗
#37 Llama 3-70B Meta AI 51% instruct-cited-nvlm official ↗
#38 Gemma 2 9B Google DeepMind 50.9% instruct-cited-phi35moe official ↗
#39 Aria Rhymes AI 50.8% from-hf-readme official ↗
#40 phi-3.5-mini Microsoft 49.8% cited-phi4-mini official ↗
#41 Pixtral 12B Mistral AI 48.1% cited-aria official ↗
#42 Llama 3.1-8B Meta AI 47.6% instruct-cited-phi35moe official ↗
#43 Llama 3.2 3B Meta AI 46.7% cited-phi4-mini official ↗
#44 Nemotron-H 8B NVIDIA 46.52% base-4-shot-cot official ↗
#45 Qwen2-7B Qwen 44.2% base-from-hf-readme official ↗
#46 Tulu 3 8B Allen Institute for AI 43.7% instruct-4-shot-cot-flex official ↗
#47 Qwen2-57B-A14B Qwen 43% base-from-hf-readme official ↗
#48 Llama 3.1-8B Meta AI 42.5% instruct-cited-tulu3 official ↗
#49 Hermes 3 70B Nous Research 41.9% cited-tulu3 official ↗
#50 Ministral 8B Mistral AI 41.8% cited-phi4-mini official ↗
#51 Yi-1.5-34B 01.AI 41.7% cited-qwen2-57b official ↗
#52 Gemini 1.5 Flash (Sep 2024) Google DeepMind 38% cited-phi35moe official ↗
#53 Qwen1.5-32B Qwen 36.1% cited-qwen2-57b official ↗
#54 Mistral NeMo Mistral AI 31.2% instruct-cited-phi35moe official ↗
#55 Mixtral 8x7B Mistral AI 30.8% cited-qwen2-57b official ↗
#56 Gemma 2 9B Google DeepMind 29.8% instruct-cited-tulu3 official ↗
#57 Gemma 7B Google DeepMind 24.3% base-cited-qwen2-7b official ↗
#58 Nous-Hermes-2-Yi-34B Nous Research 21.8% cited-nvlm official ↗
#59 Qwen2-1.5B Qwen 21.7% base-from-hf-readme official ↗
#60 Llama 3-8B Meta AI 20.5% base-cited-qwen2-7b official ↗
#61 Qwen1.5-7B Qwen 20.3% base-cited-qwen2-7b official ↗
#62 Mistral 7B Mistral AI 13.1% base-cited-qwen2-7b official ↗
#63 Gemma 2B Google DeepMind 11.8% cited-qwen2 official ↗
#64 Qwen2-0.5B Qwen 10.7% base-from-hf-readme official ↗
#65 Phi-2 Microsoft 3.5% cited-qwen2 official ↗

Frequently asked questions about MATH

What is the MATH benchmark?

12,500 high-school competition math problems with step-by-step solutions.

How is the MATH benchmark scored?

MATH is scored using the accuracy (%) metric, where a higher score is better. The maximum achievable score is 100.000. GenAIList aggregates reported scores from model providers and papers into a single ranked leaderboard.

Which AI model scores highest on MATH?

As of the latest reported scores on GenAIList, DeepSeek-V3 achieves the highest result on MATH with a score of 89.7%.

Is a higher MATH score better?

Yes. On MATH a higher score indicates better performance, so models near the top of the leaderboard are the strongest.