ARC AI2
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 77 Ergebnissen; die Tabelle führt alle auf.
| # | Modell | Organisation | Ergebnis | Nachweis |
|---|---|---|---|---|
| 1 | DeepSeek-V3 Open weights (restricted use) | DeepSeek | 95,3 % | – |
| 2 | Llama-3.1-405B Open weights (restricted use) | Meta AI | 95,3 % | – |
| 3 | Qwen2.5-72B Open weights (unrestricted) | Alibaba | 94,5 % | – |
| 4 | DeepSeek-V2 Open weights (restricted use) | DeepSeek | 92,2 % | – |
| 5 | Phi-3-medium-128k-instruct Open weights (unrestricted) | Microsoft | 91,6 % | – |
| 6 | Phi-3-small-8k-instruct Open weights (unrestricted) | Microsoft | 90,7 % | – |
| 7 | GPT-3.5 Turbo (Nov 2023) API access | OpenAI | 87,4 % | – |
| 8 | claude-instant-1.2 API access | Anthropic | 86,3 % | Protokoll oder Quelle |
| 9 | StableBeluga2 Open weights (non-commercial) | Stability AI | 86,1 % | Protokoll oder Quelle |
| 10 | claude-instant-1.1 API access | Anthropic | 85,7 % | Protokoll oder Quelle |
| 11 | text-davinci-002 API access | OpenAI | 85,2 % | – |
| 12 | Phi-3-mini-4k-instruct Open weights (unrestricted) | Microsoft | 84,9 % | – |
| 13 | Qwen-14B Open weights (restricted use) | Alibaba | 84,4 % | Protokoll oder Quelle |
| 14 | Meta-Llama-3-8B-Instruct Open weights (restricted use) | Meta AI | 82,8 % | – |
| 15 | internlm-20b | – | 81,7 % | Protokoll oder Quelle |
| 16 | phi-2 Open weights (unrestricted) | Microsoft | 75,9 % | – |
| 17 | Qwen-7B Open weights (restricted use) | Alibaba | 75,3 % | Protokoll oder Quelle |
| 18 | Qwen2.5-Coder-32B Open weights (unrestricted) | Alibaba | 70,5 % | – |
| 19 | LLaMA-65B Open weights (non-commercial) | Meta AI | 69,5 % | Protokoll oder Quelle |
| 20 | internlm-7b | – | 69,5 % | Protokoll oder Quelle |
| 21 | PaLM 2-L | – | 69,2 % | – |
| 22 | falcon-180B Open weights (restricted use) | Technology Innovation Institute | 67,8 % | – |
| 23 | LLaMA-33B Open weights (non-commercial) | Meta AI | 67,5 % | Protokoll oder Quelle |
| 24 | Qwen2.5-Coder-14B | – | 66 % | – |
| 25 | PaLM 2-M | – | 64,9 % | – |
| 26 | DeepSeek-Coder-V2-Base Open weights (restricted use) | DeepSeek | 64,3 % | – |
| 27 | chatglm2-6b | – | 61 % | Protokoll oder Quelle |
| 28 | Qwen2.5-Coder-7B Open weights (unrestricted) | Alibaba | 60,9 % | – |
| 29 | PaLM 540B Unreleased | Google Research | 60,1 % | – |
| 30 | Mixtral-8x7B-v0.1 Open weights (unrestricted) | Mistral AI | 59,7 % | – |
| 31 | PaLM 2-S | – | 59,6 % | – |
| 32 | Llama-2-70b-hf Open weights (restricted use) | Meta AI | 57,4 % | – |
| 33 | DeepSeek-Coder-V2-Lite-Base | – | 57,3 % | – |
| 34 | falcon-40b Open weights (unrestricted) | Technology Innovation Institute | 56,7 % | – |
| 35 | Yi-9B | – | 55,6 % | – |
| 36 | Mistral-7B-v0.1 Open weights (unrestricted) | Mistral AI | 55,5 % | – |
| 37 | Nemotron-4 15B Unreleased | Nvidia | 55,5 % | – |
| 38 | INTELLECT-1-Instruct | Prime Intellect,Hugging Face,Arcee AI | 54,5 % | Protokoll oder Quelle |
| 39 | Llama-2-34b Unreleased | Meta AI | 54,5 % | – |
| 40 | Qwen-1.8B | – | 53,2 % | Protokoll oder Quelle |
| 41 | gemma-7b Open weights (restricted use) | Google DeepMind | 53,2 % | – |
| 42 | Qwen2.5-Coder-3B | – | 52,9 % | – |
| 43 | LLaMA-13B Open weights (non-commercial) | Meta AI | 52,7 % | Protokoll oder Quelle |
| 44 | PaLM 62B | – | 52,5 % | – |
| 45 | text-davinci-001 API access | OpenAI | 51,4 % | – |
| 46 | mpt-30b Open weights (unrestricted) | MosaicML | 50,6 % | – |
| 47 | Yi-6B Open weights (restricted use) | 01.AI | 50,3 % | – |
| 48 | Llama-2-13b Open weights (restricted use) | Meta AI | 49,4 % | – |
| 49 | LLaMA-7B Open weights (non-commercial) | Meta AI | 47,6 % | Protokoll oder Quelle |
| 50 | starcoder2-15b Open weights (restricted use) | Hugging Face,ServiceNow,Nvidia,BigCode | 47,2 % | – |
| 51 | Llama-2-7b Open weights (restricted use) | Meta AI | 45,9 % | – |
| 52 | Qwen2.5-Coder-1.5B Open weights (unrestricted) | Alibaba | 45,2 % | – |
| 53 | falcon-7b Open weights (unrestricted) | Technology Innovation Institute | 44,5 % | – |
| 54 | phi-1_5 Open weights (unrestricted) | Microsoft | 44,4 % | – |
| 55 | vicuna-13b-v1.1 | – | 43,2 % | – |
| 56 | mpt-7b Open weights (unrestricted) | MosaicML | 42,6 % | Protokoll oder Quelle |
| 57 | deepseek-coder-33b-base Open weights (restricted use) | DeepSeek,Peking University | 42,2 % | – |
| 58 | gemma-2b Open weights (restricted use) | Google DeepMind | 42,1 % | – |
| 59 | xgen-7b-8k-base Open weights (unrestricted) | Salesforce | 41,2 % | – |
| 60 | gpt-neox-20b Open weights (unrestricted) | EleutherAI | 41,1 % | – |
| 61 | dolly-v2-12b Open weights (unrestricted) | Databricks | 39,6 % | – |
| 62 | RedPajama-INCITE-7B-Base | – | 39,1 % | – |
| 63 | open_llama_7b | – | 38,7 % | – |
| 64 | starcoder2-7b Open weights (restricted use) | Hugging Face,ServiceNow,Nvidia,BigCode | 38,7 % | – |
| 65 | Baichuan-2-13B-Base Open weights (restricted use) | Baichuan | 38 % | Protokoll oder Quelle |
| 66 | deepseek-coder-6.7b-base Open weights (restricted use) | DeepSeek,Peking University | 36,4 % | – |
| 67 | gpt-j-6b Open weights (unrestricted) | EleutherAI,LAION | 36,3 % | – |
| 68 | opt-13b | – | 35,8 % | – |
| 69 | CodeQwen1.5-7B | – | 35,7 % | – |
| 70 | Qwen2.5-Coder-0.5B | – | 34,4 % | – |
| 71 | starcoder2-3b Open weights (restricted use) | Hugging Face,ServiceNow,Nvidia,BigCode | 34,2 % | – |
| 72 | Baichuan-2-7B-Base Open weights (restricted use) | Baichuan | 32,5 % | Protokoll oder Quelle |
| 73 | Cerebras-GPT-13B Open weights (unrestricted) | Cerebras Systems | 32,4 % | – |
| 74 | stablelm-tuned-alpha-7b | – | 27 % | – |
| 75 | deepseek-coder-1.3b-base Open weights (restricted use) | DeepSeek,Peking University | 25,4 % | – |
| 76 | gpt2-xl Open weights (unrestricted) | OpenAI | 25 % | – |
| 77 | opt-1.3b Open weights (non-commercial) | Meta AI | 23,2 % | – |
