GSM8K
Grade-school maths word problems. A 2021 benchmark, saturated for current models; kept for history.
As of 19 Sept 2026, Deepseek Coder v2 leads GSM8K on BenchLeader with 94.5%, ahead of Qwen2.5-Coder-14B at 94.2%, across 74 model configurations with a published result.
- Published by
- Epoch AI Benchmarking Hub
- Category
- Maths
- Index weight
- Reference only
- Models
- 74
- Data as of
- 19 Sept 2026
CC BY 4.0 — Epoch AI, ‘AI Benchmarking Hub’, epoch.ai/benchmarks. Mirrored boards credit their original publishers.
What the test looks like
Grade-school maths word problems.
How it is scored
Accuracy, as compiled by Epoch AI from published results.
What to keep in mind
Saturated: frontier models score near the ceiling, so it separates only older or smaller models, and results are largely self-reported by developers.
- 1Deepseek Coder v294.5%
- 2Qwen2.5-Coder-14B94.2%
- 3Qwen2.5-Coder 32B Instruct93.0%
- 4GPT 4 031492.0%
- 5GPT-4o mini91.3%
- 6GPT 4 061390.0%
- 7Phi-3.5-MoE88.7%
- 8DeepSeek-Coder-V2-Lite-Instruct87.6%
- 9Qwen2.5-Coder 7B Instruct86.7%
- 10Claude Instant86.7%
- 11Phi-3.5-mini86.2%
- 12DeepSeek-Coder-V2 236B85.8%
- 13Gemma 2 9B84.9%
- 14Mistral NeMo84.2%
- 15Llama 3.1 8B82.4%
74 of 74
| # | ||||
|---|---|---|---|---|
| 1 |  Deepseek Coder v2DeepSeekopen ↗ | 94.5% | 45.4 | 2024-06-17 |
| 2 |  Qwen2.5-Coder-14BAlibaba | 94.2% | – | 2024-11-06 |
| 3 | Qwen2.5-Coder 32B InstructAlibabaopen ↗ | 93.0% | 42.6 | 2024-11-21 |
| 4 |  GPT 4 0314OpenAI | 92.0% | 43.9 | 2023-03-14 |
| 5 | GPT-4o miniOpenAI | 91.3% | 36.2 | 2024-07-18 |
| 6 | GPT 4 0613OpenAI | 90.0% | 40.2 | 2023-06-13 |
| 7 |  Phi-3.5-MoEMicrosoft | 88.7% | – | 2024-08-17 |
| 8 | DeepSeek-Coder-V2-Lite-InstructDeepSeekopen | 87.6% | – | 2024-06-13 |
| 9 | Qwen2.5-Coder 7B InstructAlibabaopen ↗ | 86.7% | – | 2024-09-17 |
| 10 |  Claude InstantAnthropic | 86.7% | – | 2023-08-09 |
| 11 | Phi-3.5-miniMicrosoft | 86.2% | – | 2024-08-16 |
| 12 | DeepSeek-Coder-V2 236BDeepSeek | 85.8% | – | 2024-06-17 |
| 13 |  Gemma 2 9BGoogle | 84.9% | 39.5 | – |
| 14 |  Mistral NeMoMistral AI | 84.2% | – | 2024-07-18 |
| 15 |  Llama 3.1 8BMetaopen | 82.4% | 36.8 | 2024-07-23 |
| 16 | Gemini 1.5 Flash 001Google | 82.4% | 36.7 | 2024-05-23 |
| 17 | Claude InstantAnthropic | 80.9% | – | – |
| 18 | Qwen2.5-Coder-3BAlibaba | 80.7% | – | 2024-11-06 |
| 19 | davinci-003OpenAI | 78.2% | – | 2022-11-28 |
| 20 |  Yi-34B01.AI | 76.0% | 35.5 | 2023-11-22 |
| 21 | Mixtral 8x7BMistral AIopen | 74.4% | 32.1 | 2023-12-11 |
| 22 | Llama 2-70BMetaopen ↗ | 69.6% | 34.9 | 2023-07-18 |
| 23 |  Stable Beluga 2Stability AI | 69.6% | – | 2023-07-20 |
| 24 | DeepSeek-Coder-V2-Lite-BaseDeepSeek | 67.1% | – | 2024-06-13 |
| 25 | Qwen2 5 Coder 1 5BAlibaba | 65.8% | – | 2024-09-18 |
| 26 |  internlm-20bShanghai AI Lab | 62.9% | – | 2023-09-18 |
| 27 | Qwen-14BAlibaba | 61.3% | – | 2023-09-28 |
| 28 | GPT-3.5 Turbo (older v0613)OpenAI | 57.8% | – | 2023-06-13 |
| 29 |  StarCoder 2 15BNVIDIA | 57.7% | – | 2024-02-20 |
| 30 | davinci-002OpenAI | 56.8% | – | 2022-03-15 |
| 31 | Palm 540BGoogle | 56.5% | – | 2022-04-04 |
| 32 | Mistral 7B v0.3Mistral AIopen | 54.4% | 32.1 | 2023-09-27 |
| 33 |  Falcon-180BTII | 54.4% | – | 2023-09-06 |
| 34 | LLaMA-65BMetaopen | 54.4% | – | 2023-02-24 |
| 35 | Falcon 2 11BTII | 53.8% | – | 2024-05-09 |
| 36 |  Baichuan2-13BBaichuan | 52.8% | – | 2023-09-06 |
| 37 | Qwen-7BAlibaba | 51.7% | – | 2023-09-28 |
| 38 | Gemma 7BGoogle | 46.4% | – | 2024-02-21 |
| 39 | Nemotron-4 15BNVIDIA | 46.0% | – | 2024-02-26 |
| 40 | Baichuan2-13BBaichuan | 45.7% | – | 2023-09-06 |
| 41 | Yi 6B01.AI | 44.9% | – | 2023-11-22 |
| 42 | LLaMA-33BMeta | 44.1% | – | 2023-02-24 |
| 43 | Llama 2-34BMeta | 42.2% | – | 2023-07-18 |
| 44 | davinci-002OpenAI | 41.5% | – | 2022-03-15 |
| 45 |  INTELLECT-1Prime Intellect | 38.6% | – | 2024-11-29 |
| 46 | CodeQwen1.5-7BUnknown | 37.7% | – | 2024-04-15 |
| 47 | Llama 2-13BMetaopen ↗ | 36.9% | 36.5 | 2023-07-18 |
| 48 | DeepSeek Coder 33BDeepSeek | 35.4% | – | 2023-11-02 |
| 49 | Qwen2.5-Coder-0.5BAlibaba | 34.5% | – | 2024-09-18 |
| 50 |  MPT-30BMosaicML | 34.4% | – | 2023-06-22 |
| 51 | Falcon-40BTII | 33.8% | – | 2023-05-25 |
| 52 | PaLM 62BGoogle | 33.0% | – | 2022-04-04 |
| 53 | StarCoder 2 7BNVIDIA | 32.7% | – | 2024-02-20 |
| 54 |  chatglm2-6bZhipu AI | 32.4% | – | 2023-06-24 |
| 55 | internlm-7bShanghai AI Lab | 31.2% | – | 2023-07-05 |
| 56 |  vicuna-13b-v1.1LMSYS | 28.1% | – | 2023-04-12 |
| 57 | Baichuan 1-13BBaichuan | 26.8% | – | 2023-07-11 |
| 58 | Baichuan 2-7BBaichuan | 24.6% | – | 2023-09-20 |
| 59 | StarCoder 2 3BNVIDIA | 21.6% | – | 2024-02-22 |
| 60 | DeepSeek Coder 6.7BDeepSeek | 21.3% | – | 2023-11-02 |
Cite as: BenchLeader, “GSM8K leaderboard”, https://www.benchleader.com/benchmarks/gsm8k, data as of 19 Sept 2026.
GSM8K: questions
- What does GSM8K measure?
- Grade-school maths word problems. Scores are reported in percent of tasks solved; higher is better.
- Which AI model leads GSM8K?
- Deepseek Coder v2 leads GSM8K with 94.5% as of 19 Sept 2026, ahead of Qwen2.5-Coder-14B at 94.2%.
- How many models have GSM8K results?
- 74 model configurations have a GSM8K result on BenchLeader, all taken from Epoch AI Benchmarking Hub.
- Who runs GSM8K and how often is it updated?
- GSM8K is published by Epoch AI Benchmarking Hub. BenchLeader re-reads the published results every morning and records the date each result was published.
- Does GSM8K count toward the BenchLeader Index?
- No. GSM8K is shown for reference but left out of the composite index.