Reasoning
WinoGrande Leaderboard
WinoGrande is a large-scale commonsense-reasoning test built around pronoun resolution: the model must pick which of two nouns a pronoun refers to in a sentence, using real-world knowledge to disambiguate. It was designed by AI2 to resist the shortcuts that let earlier models game its predecessor, Winograd.
Source: epoch46 open models ranked+34 proprietaryData through Dec 2024
All models ranked on WinoGrande
Proprietary / closed models are shown dimmed — you can't run them locally, but they show where the open field stands.
| # | Model | Score |
|---|---|---|
| 1 | Llama 3.1 405B · 405.9B | 89.2% |
| 2 | Claude 3 Opus (Feb 29, 2024) · proprietary | 88.5% |
| 3 | GPT 4 (Mar 14) · proprietary | 87.5% |
| 4 | GPT 4 32K (Mar 14) · proprietary | 87.5% |
| 5 | Falcon 180B · 180B | 87.1% |
| 6 | DeepSeek v2 · 235.7B | 86.3% |
| 7 | DeepSeek v3 · 684.5B | 85.2% |
| 8 | PaLM 540B · proprietary | 85.1% |
| 9 | DeepSeek Coder V2 Base · proprietary | 83.7% |
| 10 | Meta Llama 3 70B · 70.6B | 83.5% |
| 11 | PaLM 2 L · proprietary | 83.0% |
| 12 | Qwen2.5 72B · 72.7B | 82.3% |
| 13 | Llama 3.1 405B Instruct · 405.9B | 82.2% |
| 14 | GPT 3.5 Turbo (Jun 13) · proprietary | 81.6% |
| 15 | Text Davinci 002 · proprietary | 81.6% |
| 16 | Phi 3 Medium 128K Instruct · proprietary | 81.5% |
| 17 | Phi 3 Small 8k Instruct · 7.4B | 81.5% |
| 18 | Qwen2.5 Coder 32B · 32.8B | 80.8% |
| 19 | Llama 2 70B HF · 69.0B | 80.2% |
| 20 | GLaM (MoE) · proprietary | 79.2% |
| 21 | PaLM 2 M · proprietary | 79.2% |
| 22 | Gemma 7B · 8.5B | 79.0% |
| 23 | Megatron Turing NLG 530B · proprietary | 78.9% |
| 24 | Falcon 11B · 11.1B | 78.3% |
| 25 | Nemotron 4 15B · proprietary | 78.0% |
| 26 | PaLM 2 S · proprietary | 77.9% |
| 27 | Text Davinci 001 · proprietary | 77.7% |
| 28 | Mixtral 8x7B v0.1 · 46.7B | 77.2% |
| 29 | Llama 65B · proprietary | 77.0% |
| 30 | PaLM 62B · proprietary | 77.0% |
| 31 | Falcon 40B · 41.8B | 76.9% |
| 32 | Qwen2.5 Coder 14B · 14.8B | 76.8% |
| 33 | Llama 2 34B · proprietary | 76.7% |
| 34 | Llama 33B · proprietary | 76.0% |
| 35 | Meta Llama 3 8B · 8.0B | 75.7% |
| 36 | Mistral 7B v0.1 · 7B | 75.3% |
| 37 | Claude 3 Sonnet (Feb 29, 2024) · proprietary | 75.1% |
| 38 | Chinchilla (70B) · proprietary | 74.9% |
| 39 | Claude 3 Haiku (Mar 07, 2024) · proprietary | 74.2% |
| 40 | Phi 1 5 · 1.4B | 73.4% |
| 41 | Llama 13B · proprietary | 73.0% |
| 42 | Yi 9B · 8.8B | 73.0% |
| 43 | DeepSeek Coder v2 Lite Base · 15.7B | 72.9% |
| 44 | Qwen2.5 Coder 7B · 7.6B | 72.9% |
| 45 | Llama 2 13B HF · 13.0B | 72.8% |
| 46 | Yi 6B · 6.1B | 71.3% |
| 47 | Mpt 30B · proprietary | 71.0% |
| 48 | Phi 3 Mini 4k Instruct · 3.8B | 70.8% |
| 49 | Vicuna 13B v1.1 · proprietary | 70.8% |
| 50 | Gopher (280B) · proprietary | 70.2% |
| 51 | Llama 7B · 6.7B | 70.1% |
| 52 | Llama 2 7B HF · 6.7B | 69.2% |
| 53 | GPT 3.5 Turbo (Nov 06) · proprietary | 68.8% |
| 54 | Mpt 7B · proprietary | 68.6% |
| 55 | Qwen2.5 Coder 3B · 3.1B | 67.4% |
| 56 | Falcon 7B · 7.2B | 67.2% |
| 57 | Open Llama 7B · proprietary | 67.0% |
| 58 | GPT Neox 20B · 20.7B | 66.1% |
| 59 | INTELLECT 1 Instruct · 10.2B | 65.8% |
| 60 | Gemma 2B · 2.5B | 65.4% |
| 61 | Meta Llama 3 8B Instruct · 8.0B | 65.0% |
| 62 | Xgen 7B 8k Base · 7B | 64.9% |
| 63 | Opt 13B · proprietary | 64.7% |
| 64 | GPT J 6B · 6B | 64.5% |
| 65 | Starcoder2 15B · 16.0B | 64.3% |
| 66 | RedPajama INCITE 7B Base · proprietary | 63.8% |
| 67 | DeepSeek Coder 33B Base · proprietary | 62.0% |
| 68 | Dolly v2 12B · proprietary | 61.8% |
| 69 | Opt 1.3b · proprietary | 61.0% |
| 70 | Cerebras GPT 13B · 13B | 60.8% |
| 71 | Qwen2.5 Coder 1.5B · 1.5B | 60.7% |
| 72 | CodeQwen1.5 7B · 7.3B | 59.8% |
| 73 | Gpt2 Xl · 1.6B | 58.3% |
| 74 | Deepseek Coder 6.7B Base · 6.7B | 57.6% |
| 75 | Starcoder2 3B · 3.0B | 57.1% |
| 76 | Starcoder2 7B · 7.2B | 57.1% |
| 77 | Qwen2.5 Coder 0.5B · 494M | 54.8% |
| 78 | Phi 2 · 2.8B | 54.7% |
| 79 | Deepseek Coder 1.3B Base · 1.3B | 53.3% |
| 80 | Stablelm Tuned Alpha 7B · 7B | 51.5% |
Score vs model size
Which models give the most quality for their size — the ones worth running locally.
- Qwen2.5 Coder 0.5B, 494M, score 54.8% — on the efficiency frontier (best score at its size or smaller).
- Phi 1 5, 1B, score 73.4% — on the efficiency frontier (best score at its size or smaller).
- Mistral 7B v0.1, 7B, score 75.3% — on the efficiency frontier (best score at its size or smaller).
- Phi 3 Small 8k Instruct, 7B, score 81.5% — on the efficiency frontier (best score at its size or smaller).
- Meta Llama 3 70B, 71B, score 83.5% — on the efficiency frontier (best score at its size or smaller).
- Falcon 180B, 180B, score 87.1% — on the efficiency frontier (best score at its size or smaller).
- Llama 3.1 405B, 406B, score 89.2% — on the efficiency frontier (best score at its size or smaller).
WinoGrande: frequently asked questions
- What is the best open LLM on WinoGrande?
- Llama 3.1 405B is the top open model on WinoGrande, scoring 89.2%. Among all models tested — including proprietary ones — it ranks #1. That puts it ahead of every proprietary model we track, including Claude 3 Opus (Feb 29, 2024) (Anthropic) at 88.5%.
- What's the best WinoGrande model you can run on a 24 GB GPU?
- Phi 3 Small 8k Instruct is the highest-scoring open model that fits in 24 GB at 4-bit quantization (about 4 GB), scoring 81.5% on WinoGrande.
- What's the best WinoGrande model you can run on a 12 GB GPU?
- Phi 3 Small 8k Instruct is the highest-scoring open model that fits in 12 GB at 4-bit quantization (about 4 GB), scoring 81.5% on WinoGrande.
- Can open models match proprietary models on WinoGrande?
- Yes — the best open model (Llama 3.1 405B, 89.2%) matches or beats every proprietary model we track on WinoGrande.
Scores aggregated from epoch. llmrun does not run this benchmark — see the source for methodology, or the about benchmarks for what it measures.