GSM8K
Das Ergebnis jedes Modells, so, wie der Benchmarking Hub von Epoch AI es veröffentlicht, von Epoch gemessen oder aus der genannten Quelle gesammelt. Facthory hat sie nicht selbst gemessen.
Daten: Epoch AI, „Capabilities & benchmarking“, epoch.ai/benchmarks, lizenziert unter CC BY 4.0. Gelesen am 16.11.2026.
Das Diagramm zeigt die 25 besten von 93 Ergebnissen; die Tabelle führt alle auf.
| # | Modell | Organisation | Ergebnis | Nachweis |
|---|---|---|---|---|
| 1 | DeepSeek-Coder-V2-Instruct Open weights (restricted use) | DeepSeek | 94,5 % | – |
| 2 | Qwen2.5-Coder-14B-Instruct | – | 94,2 % | – |
| 3 | Qwen2.5-Coder-32B-Instruct Open weights (unrestricted) | Alibaba | 93 % | – |
| 4 | GPT-4 (Mar 2023) API access | OpenAI | 92 % | – |
| 5 | gpt-4o-mini-2024-07-18 API access | OpenAI | 91,3 % | – |
| 6 | Qwen2.5-Coder-32B Open weights (unrestricted) | Alibaba | 91,1 % | – |
| 7 | GPT-4 (Jun 2023) API access | OpenAI | 90 % | – |
| 8 | Phi-3.5-MoE-instruct Open weights (unrestricted) | Microsoft | 88,7 % | – |
| 9 | Qwen2.5-Coder-14B | – | 88,7 % | – |
| 10 | DeepSeek-Coder-V2-Lite-Instruct | – | 87,6 % | – |
| 11 | Qwen2.5-Coder-7B-Instruct Open weights (unrestricted) | Alibaba | 86,7 % | – |
| 12 | claude-instant-1.2 API access | Anthropic | 86,7 % | Protokoll oder Quelle |
| 13 | Phi-3.5-mini-instruct Open weights (unrestricted) | Microsoft | 86,2 % | – |
| 14 | DeepSeek-Coder-V2-Base Open weights (restricted use) | DeepSeek | 85,8 % | – |
| 15 | gemma-2-9b Open weights (restricted use) | Google DeepMind | 84,9 % | – |
| 16 | Mistral-Nemo-Base-2407 Open weights (unrestricted) | Mistral AI | 84,2 % | – |
| 17 | Qwen2.5-Coder-7B Open weights (unrestricted) | Alibaba | 83,9 % | – |
| 18 | Llama-3.1-8B-Instruct Open weights (restricted use) | Meta AI | 82,4 % | – |
| 19 | Gemini 1.5 Flash (May 2024) API access | Google DeepMind | 82,4 % | – |
| 20 | claude-instant-1.1 API access | Anthropic | 80,9 % | Protokoll oder Quelle |
| 21 | Qwen2.5-Coder-3B-Instruct | – | 80,7 % | – |
| 22 | Qwen2.5-Coder-3B | – | 75,7 % | – |
| 23 | Mixtral-8x7B-v0.1 Open weights (unrestricted) | Mistral AI | 74,4 % | – |
| 24 | Yi-34B-Chat Open weights (restricted use) | 01.AI | 71,7 % | – |
| 25 | StableBeluga2 Open weights (non-commercial) | Stability AI | 69,6 % | Protokoll oder Quelle |
| 26 | Yi-34B Open weights (restricted use) | 01.AI | 67,2 % | – |
| 27 | DeepSeek-Coder-V2-Lite-Base | – | 67,1 % | – |
| 28 | Qwen2.5-Coder-1.5B Open weights (unrestricted) | Alibaba | 65,8 % | – |
| 29 | internlm-20b | – | 62,9 % | – |
| 30 | Qwen-14B Open weights (restricted use) | Alibaba | 60,1 % | – |
| 31 | Qwen-14B-Chat Open weights (restricted use) | Alibaba | 59,5 % | – |
| 32 | starcoder2-15b Open weights (restricted use) | Hugging Face,ServiceNow,Nvidia,BigCode | 57,7 % | – |
| 33 | code-davinci-002 API access | OpenAI | 56,8 % | Protokoll oder Quelle |
| 34 | PaLM 540B Unreleased | Google Research | 56,5 % | – |
| 35 | falcon-180B Open weights (restricted use) | Technology Innovation Institute | 54,4 % | – |
| 36 | Falcon 2-11B Open weights (restricted use) | Technology Innovation Institute | 53,8 % | – |
| 37 | Baichuan-2-13B-Base Open weights (restricted use) | Baichuan | 52,8 % | – |
| 38 | Qwen-7B Open weights (restricted use) | Alibaba | 51,7 % | – |
| 39 | text-davinci-003 | OpenAI | 50,6 % | Protokoll oder Quelle |
| 40 | Llama-2-70b-hf Open weights (restricted use) | Meta AI | 48,4 % | Protokoll oder Quelle |
| 41 | Llama-2-70b-chat Open weights (restricted use) | Meta AI | 47,1 % | – |
| 42 | GPT-3.5 Turbo (June 2023) API access | OpenAI | 46,9 % | Protokoll oder Quelle |
| 43 | LLaMA-65B Open weights (non-commercial) | Meta AI | 46,6 % | Protokoll oder Quelle |
| 44 | gemma-7b Open weights (restricted use) | Google DeepMind | 46,4 % | – |
| 45 | Nemotron-4 15B Unreleased | Nvidia | 46 % | – |
| 46 | Baichuan2-13B-Chat | – | 45,7 % | – |
| 47 | Llama-2-34b Unreleased | Meta AI | 42,2 % | – |
| 48 | text-davinci-002 API access | OpenAI | 41,5 % | Protokoll oder Quelle |
| 49 | INTELLECT-1-Instruct | Prime Intellect,Hugging Face,Arcee AI | 38,6 % | Protokoll oder Quelle |
| 50 | Yi-6B-Chat Open weights (restricted use) | 01.AI | 38,4 % | – |
| 51 | Mistral-7B-v0.1 Open weights (unrestricted) | Mistral AI | 38,1 % | Protokoll oder Quelle |
| 52 | CodeQwen1.5-7B | – | 37,7 % | – |
| 53 | Llama-2-13b-chat Open weights (restricted use) | Meta AI | 36,9 % | – |
| 54 | Mistral-7B-Instruct-v0.2 Open weights (unrestricted) | Mistral AI | 35,4 % | – |
| 55 | deepseek-coder-33b-base Open weights (restricted use) | DeepSeek,Peking University | 35,4 % | – |
| 56 | Qwen2.5-Coder-0.5B | – | 34,5 % | – |
| 57 | mpt-30b-instruct Open weights (unrestricted) | MosaicML | 34,4 % | Protokoll oder Quelle |
| 58 | falcon-40b-instruct Open weights (unrestricted) | Technology Innovation Institute | 33,8 % | Protokoll oder Quelle |
| 59 | PaLM 62B | – | 33 % | – |
| 60 | starcoder2-7b Open weights (restricted use) | Hugging Face,ServiceNow,Nvidia,BigCode | 32,7 % | – |
| 61 | Yi-6B Open weights (restricted use) | 01.AI | 32,5 % | – |
| 62 | LLaMA-33B Open weights (non-commercial) | Meta AI | 32 % | Protokoll oder Quelle |
| 63 | internlm-7b | – | 31,2 % | Protokoll oder Quelle |
| 64 | chatglm2-6b | – | 28,9 % | – |
| 65 | vicuna-13b-v1.1 | – | 28,1 % | – |
| 66 | Baichuan-13B-Base Open weights (restricted use) | Baichuan | 26,8 % | – |
| 67 | falcon-40b Open weights (unrestricted) | Technology Innovation Institute | 25 % | Protokoll oder Quelle |
| 68 | Baichuan-2-7B-Base Open weights (restricted use) | Baichuan | 24,5 % | – |
| 69 | Llama-2-13b Open weights (restricted use) | Meta AI | 24,5 % | Protokoll oder Quelle |
| 70 | vicuna-13b-v1.3 Open weights (restricted use) | Large Model Systems Organization,University of California (UC) Berkeley | 22,6 % | Protokoll oder Quelle |
| 71 | starcoder2-3b Open weights (restricted use) | Hugging Face,ServiceNow,Nvidia,BigCode | 21,6 % | – |
| 72 | deepseek-coder-6.7b-base Open weights (restricted use) | DeepSeek,Peking University | 21,3 % | – |
| 73 | Qwen-1.8B | – | 21,2 % | Protokoll oder Quelle |
| 74 | gemma-2b Open weights (restricted use) | Google DeepMind | 17,7 % | – |
| 75 | mpt-30b Open weights (unrestricted) | MosaicML | 16,4 % | Protokoll oder Quelle |
| 76 | internlm-chat-20b | – | 15,7 % | – |
| 77 | LLaMA-13B Open weights (non-commercial) | Meta AI | 15,4 % | Protokoll oder Quelle |
| 78 | Llama-2-7b Open weights (restricted use) | Meta AI | 14,6 % | – |
| 79 | bloom Open weights (restricted use) | Hugging Face,BigScience | 9,5 % | Protokoll oder Quelle |
| 80 | Baichuan-7B Open weights (non-commercial) | Baichuan | 9,2 % | – |
| 81 | davinci API access | OpenAI | 9 % | Protokoll oder Quelle |
| 82 | mpt-7b Open weights (unrestricted) | MosaicML | 8,6 % | – |
| 83 | LLaMA-7B Open weights (non-commercial) | Meta AI | 8 % | Protokoll oder Quelle |
| 84 | deepseek-coder-1.3b-base Open weights (restricted use) | DeepSeek,Peking University | 4,4 % | – |
| 85 | falcon-7b Open weights (unrestricted) | Technology Innovation Institute | 4 % | Protokoll oder Quelle |
| 86 | opt-175b Open weights (non-commercial) | Meta AI | 4 % | Protokoll oder Quelle |
| 87 | opt-66b Open weights (non-commercial) | Meta AI | 1,8 % | Protokoll oder Quelle |
| 88 | curie | OpenAI | 1,6 % | Protokoll oder Quelle |
| 89 | babbage | OpenAI | 0,7 % | Protokoll oder Quelle |
| 90 | ada | OpenAI | 0,6 % | Protokoll oder Quelle |
| 91 | text-curie-001 API access | OpenAI | 0,6 % | Protokoll oder Quelle |
| 92 | text-ada-001 | OpenAI | 0,4 % | Protokoll oder Quelle |
| 93 | text-babbage-001 API access | OpenAI | 0 % | Protokoll oder Quelle |
