🔌
New: connect Claude & other AIs to GenAIList over MCP — research the catalog and contribute to the shared knowledge base. Learn how →
// BENCHMARK
GSM8K benchmark
AI model leaderboard for the GSM8K benchmark. Compare how large language models score on GSM8K, see the full ranking, and understand what this AI benchmark measures. Claude 3.5 Sonnet currently leads with 96.9%. 8,500 grade-school word problems requiring multi-step arithmetic reasoning.
Leaderboard
| # | Model | Organization | Score | Variant | Source |
|---|---|---|---|---|---|
| #1 | Claude 3.5 Sonnet | Anthropic | 96.9% | 1022-cited-minimax-text | official ↗ |
| #2 | Llama 3.1-405B | Meta AI | 96.8% | instruct-cited-nvlm | official ↗ |
| #3 | Llama 3.1-405B | Meta AI | 96.7% | instruct-cited-minimax-text | official ↗ |
| #4 | DeepSeek-V3 | DeepSeek | 96.7% | cited-minimax-text | official ↗ |
| #5 | Claude 3.5 Sonnet | Anthropic | 96.4% | cited-nvlm | official ↗ |
| #6 | Qwen2.5 Instruct (72B) | Qwen | 95.8% | instruct-cited-minimax-text | official ↗ |
| #7 | GPT-4o (Nov 2024) | OpenAI | 95.6% | cited-minimax-text | official ↗ |
| #8 | Tulu 3 405B | Allen Institute for AI | 95.5% | instruct-8-shot-cot | official ↗ |
| #9 | Llama 3.1-405B | Meta AI | 95.4% | instruct-cited-tulu3 | official ↗ |
| #10 | Gemini 2.0 Flash | Google DeepMind | 95.4% | exp-cited-minimax-text | official ↗ |
| #11 | Gemini 1.5 Pro | Google DeepMind | 95.2% | 002-cited-minimax-text | official ↗ |
| #12 | Llama 3.1-70B | Meta AI | 95.1% | instruct-cited-nvlm | official ↗ |
| #13 | Llama 3.1-70B | Meta AI | 95.1% | cited-telechat2 | official ↗ |
| #14 | MiniMax-Text-01 | MiniMax | 94.8% | from-hf-readme | official ↗ |
| #15 | DeepSeek-V3 | DeepSeek | 94.1% | cited-tulu3 | official ↗ |
| #16 | Nemotron-H 56B | NVIDIA | 93.71% | base-8-shot-cot | official ↗ |
| #17 | Llama 3.1-70B | Meta AI | 93.7% | instruct-cited-tulu3 | official ↗ |
| #18 | Nemotron-H 47B | NVIDIA | 93.3% | base-8-shot-cot | official ↗ |
| #19 | Llama 3-70B | Meta AI | 93% | instruct-cited-nvlm | official ↗ |
| #20 | NVLM-D 72B | NVIDIA | 92.9% | megatron-from-hf-readme | official ↗ |
| #21 | Hermes 3 405B | Nous Research | 92.7% | cited-tulu3 | official ↗ |
| #22 | Nemotron-4 340B | NVIDIA | 92.3% | 0-shot-instruct-from-hf-readme | official ↗ |
| #23 | DeepSeek-V2 (MoE-236B) | DeepSeek | 92.2% | cited-telechat2 | official ↗ |
| #24 | Telechat2-115B | China Telecom | 92.2% | from-hf-readme | official ↗ |
| #25 | GPT-4o (Nov 2024) | OpenAI | 91.7% | cited-tulu3 | official ↗ |
| #26 | GPT-4o mini | OpenAI | 91.3% | cited-phi35moe | official ↗ |
| #27 | Qwen-72B | Qwen | 91.1% | instruct-cited-nvlm | official ↗ |
| #28 | Qwen2-72B | Qwen | 91.1% | instruct-cited-telechat2 | official ↗ |
| #29 | TeleChat2-35B | China Telecom | 91% | from-hf-readme | official ↗ |
| #30 | Gemini 1.5 Pro | Google DeepMind | 90.8% | aug-2024-cited-nvlm | official ↗ |
| #31 | Hermes 3 70B | Nous Research | 90% | cited-tulu3 | official ↗ |
| #32 | Qwen2.5 Instruct (72B) | Qwen | 89.5% | instruct-cited-tulu3 | official ↗ |
| #33 | Qwen2.5-7B | Qwen | 88.7% | cited-phi4-mini | official ↗ |
| #34 | Phi-3.5-MoE | Microsoft | 88.7% | instruct-8-shot-cot | official ↗ |
| #35 | Phi-4 Mini | Microsoft | 88.6% | instruct-8-shot-cot | official ↗ |
| #36 | Tulu 3 8B | Allen Institute for AI | 87.6% | instruct-8-shot-cot | official ↗ |
| #37 | Granite-4.0-H-Small | IBM | 87.27% | 8-shot | official ↗ |
| #38 | Nemotron-H 8B | NVIDIA | 87.11% | base-8-shot-cot | official ↗ |
| #39 | TeleChat2-7B | China Telecom | 86.8% | from-hf-readme | official ↗ |
| #40 | Qwen1.5-110B | Qwen | 85.4% | cited-telechat2 | official ↗ |
| #41 | Gemma 2 9B | Google DeepMind | 84.9% | instruct-cited-phi35moe | official ↗ |
| #42 | Granite-4.0-H-Tiny | IBM | 84.69% | 8-shot | official ↗ |
| #43 | Qwen2.5-7B | Qwen | 84.46% | instruct-cited-granite-3-2 | official ↗ |
| #44 | Mistral NeMo | Mistral AI | 84.2% | instruct-cited-phi35moe | official ↗ |
| #45 | Llama 3.1-8B | Meta AI | 83.4% | instruct-cited-tulu3 | official ↗ |
| #46 | Llama 3.1-8B | Meta AI | 83.24% | instruct-cited-granite-3-2 | official ↗ |
| #47 | Yi-1.5-34B | 01.AI | 82.7% | cited-qwen2-57b | official ↗ |
| #48 | Llama 3.1-8B | Meta AI | 82.6% | instruct-5-shot-cited-falcon3 | official ↗ |
| #49 | Gemini 1.5 Flash (Sep 2024) | Google DeepMind | 82.4% | cited-phi35moe | official ↗ |
| #50 | Granite 3.2 8B | IBM | 81.65% | instruct-from-hf-readme | official ↗ |
| #51 | Qwen2.5-7B | Qwen | 81.5% | base-cited-olmo2 | official ↗ |
| #52 | Falcon3-7B | Technology Innovation Institute | 81.4% | instruct-5-shot | official ↗ |
| #53 | Granite-4.0-H-Micro | IBM | 81.35% | 8-shot | official ↗ |
| #54 | Qwen2-57B-A14B | Qwen | 80.7% | base-from-hf-readme | official ↗ |
| #55 | Qwen2.5-3B | Qwen | 80.6% | cited-phi4-mini | official ↗ |
| #56 | Ministral 8B | Mistral AI | 80.1% | cited-phi4-mini | official ↗ |
| #57 | Qwen2-7B | Qwen | 79.9% | base-from-hf-readme | official ↗ |
| #58 | Gemma 2 9B | Google DeepMind | 79.7% | instruct-cited-tulu3 | official ↗ |
| #59 | Granite 3.1 8B | IBM | 79.15% | instruct-cited-granite-3-2 | official ↗ |
| #60 | Nous-Hermes-2-Yi-34B | Nous Research | 78.6% | cited-nvlm | official ↗ |
| #61 | OLMo 2 32B | Allen Institute for AI | 78.5% | base-from-hf-readme | official ↗ |
| #62 | DeepSeek-R1-Distill-Qwen-7B | Qwen | 78.47% | cited-granite-3-2 | official ↗ |
| #63 | DeepSeek-R1-Distill-Qwen-7B | DeepSeek | 78.47% | cited-granite-3-2 | official ↗ |
| #64 | phi-3.5-mini | Microsoft | 76.9% | cited-phi4-mini | official ↗ |
| #65 | Qwen1.5-32B | Qwen | 76.8% | cited-qwen2-57b | official ↗ |
| #66 | Llama 3.2 3B | Meta AI | 75.6% | cited-phi4-mini | official ↗ |
| #67 | DeepSeek-R1-Distill-Llama-8B | DeepSeek | 72.18% | cited-granite-3-2 | official ↗ |
| #68 | Qwen2.5-7B | Qwen | 72% | instruct-5-shot-cited-falcon3 | official ↗ |
| #69 | Gemma 2 9B | Google DeepMind | 70.1% | base-cited-olmo2 | official ↗ |
| #70 | Granite 3.1 2B | IBM | 67.55% | instruct-cited-granite-3-2 | official ↗ |
| #71 | Granite 3.2 2B | IBM | 67.02% | instruct-from-hf-readme | official ↗ |
| #72 | TeleChat2-3B | China Telecom | 64.7% | from-hf-readme | official ↗ |
| #73 | Qwen1.5-7B | Qwen | 62.5% | base-cited-qwen2-7b | official ↗ |
| #74 | Mixtral 8x7B | Mistral AI | 62.5% | cited-qwen2-57b | official ↗ |
| #75 | Jamba | AI21 Labs | 59.9% | cited-qwen2-57b | official ↗ |
| #76 | Qwen2-1.5B | Qwen | 58.5% | base-from-hf-readme | official ↗ |
| #77 | Phi-2 | Microsoft | 57.2% | cited-qwen2 | official ↗ |
| #78 | Llama 3.1-8B | Meta AI | 56.5% | base-cited-olmo2 | official ↗ |
| #79 | Llama 3-8B | Meta AI | 56% | base-cited-qwen2-7b | official ↗ |
| #80 | Falcon Mamba | Technology Innovation Institute | 52.54% | base-from-hf-readme | official ↗ |
| #81 | Mistral 7B | Mistral AI | 52.2% | base-cited-qwen2-7b | official ↗ |
| #82 | Gemma 7B | Google DeepMind | 50.87% | cited-falcon-mamba | official ↗ |
| #83 | Llama 3.1-8B | Meta AI | 47.92% | base-cited-falcon-mamba | official ↗ |
| #84 | Gemma 7B | Google DeepMind | 46.4% | base-cited-qwen2-7b | official ↗ |
| #85 | Llama 3-8B | Meta AI | 45.19% | base-cited-falcon-mamba | official ↗ |
| #86 | Mistral 7B | Mistral AI | 37.83% | v0-1-cited-falcon-mamba | official ↗ |
| #87 | Qwen2-0.5B | Qwen | 36.5% | base-from-hf-readme | official ↗ |
| #88 | Llama 2-13B | Meta AI | 28.1% | base-cited-olmo2 | official ↗ |
| #89 | Gemma 2B | Google DeepMind | 17.7% | cited-qwen2 | official ↗ |
| #90 | DCLM 7B | Apple | 2.5% | cot-from-hf-readme | official ↗ |
Frequently asked questions about GSM8K
What is the GSM8K benchmark?
8,500 grade-school word problems requiring multi-step arithmetic reasoning.
How is the GSM8K benchmark scored?
GSM8K 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 GSM8K?
As of the latest reported scores on GenAIList, Claude 3.5 Sonnet achieves the highest result on GSM8K with a score of 96.9%.
Is a higher GSM8K score better?
Yes. On GSM8K a higher score indicates better performance, so models near the top of the leaderboard are the strongest.