Winogrande
ReportedFill-in-the-blank sentences where the right answer needs common sense about who or what a word refers to.
- Kind
- Published by a third party's leaderboard, collected by Epoch AI. The leaderboard is named and linked.
- Score shown
- Epoch's column "Accuracy", as a percentage, higher is better; random guessing scores about 50.0%. The source gives no error bars.
- Results
- 137 in all: 85 for open-weight models, 43 for closed ones, 9 where Epoch does not say. 36 linked to a model page here.
Best open-weight result: Llama-3.1-405B, 89.2%, number 1 of 137 results.
| # | Model | Weights | Setting | Score | Date | Source |
|---|---|---|---|---|---|---|
| 1 | Llama-3.1-405BMeta AIno page match | open | Shots: 5 | 89.2% | not dated | DeepSeek-V3 Technical Report |
| 2 | claude-3-opus-20240229Anthropic | closed | Shots: 5 | 88.5% | not dated | The Claude 3 Model Family: Opus, Sonnet, Haiku |
| 3 | gpt-4-0314OpenAI | closed | Shots: 5 | 87.5% | not dated | GPT-4 technical report |
| 4 | gpt-4-32k-0314OpenAI | closed | Shots: 5 | 87.5% | not dated | The Falcon Series of Open Language Models |
| 5 | falcon-180BTechnology Innovation Instituteno page match | open | Shots: 5 | 87.1% | not dated | The Falcon Series of Open Language Models |
| 6 | DeepSeek-V2DeepSeekno page match | open | Shots: 5 | 86.3% | not dated | DeepSeek-V3 Technical Report |
| 7 | deepseek-ai/DeepSeek-V3Epoch: DeepSeek-V3, DeepSeek | open | Shots: 5 | 85.2% | not dated | DeepSeek-V3 Technical Report |
| 8 | PaLM 540BGoogle Research | closed | Shots: 5 | 85.1% | not dated | GPT-4 Technical report |
| 9 | PaLM 540BGoogle Research | closed | Shots: few | 85.1% | not dated | PaLM: Scaling Language Modeling with Pathways |
| 10 | deepseek-ai/DeepSeek-Coder-V2-BaseEpoch: DeepSeek-Coder-V2-Base, DeepSeek | open | default | 83.7% | not dated | Qwen2.5-Coder Technical Report |
| 11 | PaLM 540BGoogle Research | closed | Shots: 1 | 83.7% | not dated | PaLM: Scaling Language Modeling with Pathways |
| 12 | Meta-Llama-3-70BMeta AIno page match | open | Shots: 5 | 83.5% | not dated | The Llama 3 Herd of Models |
| 13 | PaLM 2-L | not stated | Shots: 1 | 83.0% | not dated | PaLM 2 Technical Report |
| 14 | Qwen2.5-72BAlibabano page match | open | Shots: 5 | 82.3% | not dated | DeepSeek-V3 Technical Report |
| 15 | meta-llama/Llama-3.1-405B-InstructEpoch: Llama-3.1-405B-Instruct, Meta AI | open | Shots: 5 | 82.2% | not dated | The Llama 3 Herd of Models |
| 16 | gpt-3.5-turbo-0613OpenAI | closed | Shots: 5 | 81.6% | not dated | The Falcon Series of Open Language Models |
| 17 | text-davinci-002OpenAI | closed | Shots: 5 | 81.6% | not dated | GPT-4 technical report |
| 18 | microsoft/Phi-3-medium-128k-instructEpoch: Phi-3-medium-128k-instruct, Microsoft | open | Shots: 5 | 81.5% | not dated | Phi-3 Technical Report: A Highly Capable Language Model Locally on Your Phone |
| 19 | microsoft/Phi-3-small-8k-instructEpoch: Phi-3-small-8k-instruct, Microsoft | open | Shots: 5 | 81.5% | not dated | Phi-3 Technical Report: A Highly Capable Language Model Locally on Your Phone |
| 20 | PaLM 540BGoogle Research | closed | Shots: 0 | 81.1% | not dated | PaLM: Scaling Language Modeling with Pathways |
| 21 | PaLM 540BGoogle Research | closed | Shots: 0 | 81.1% | not dated | LLaMA: Open and Efficient Foundation Language Models |
| 22 | Qwen2.5-Coder-32BAlibabano page match | open | default | 80.8% | not dated | Qwen2.5-Coder Technical Report |
| 23 | meta-llama/Llama-2-70b-hfEpoch: Llama-2-70b-hf, Meta AI | open | default | 80.2% | not dated | Llama 2: Open Foundation and Fine-Tuned Chat Models |
| 24 | GLaM (MoE)Google | closed | Shots: 8 | 79.2% | not dated | GLaM: Efficient Scaling of Language Models with Mixture-of-Experts |
| 25 | PaLM 2-M | not stated | Shots: 1 | 79.2% | not dated | PaLM 2 Technical Report |
| 26 | gemma-7bGoogle DeepMindno page match | open | default | 79.0% | not dated | Gemma: Open Models Based on Gemini Research and Technology |
| 27 | Megatron-Turing NLG 530BMicrosoft,Nvidia | closed | default | 78.9% | not dated | GLaM: Efficient Scaling of Language Models with Mixture-of-Experts |
| 28 | tiiuae/falcon-11BEpoch: falcon-11b, Technology Innovation Institute | open | Shots: 5 | 78.3% | not dated | Falcon2-11B Technical Report |
| 29 | tiiuae/falcon-11BEpoch: falcon-11b, Technology Innovation Institute | open | Shots: 5 | 78.3% | not dated | Falcon2-11B Technical Report |
| 30 | Nemotron-4 15BNvidia | closed | Shots: 0 | 78.0% | not dated | Nemotron-4 15B Technical Report |
| 31 | PaLM 2-S | not stated | Shots: 1 | 77.9% | not dated | PaLM 2 Technical Report |
| 32 | text-davinci-001OpenAI | closed | Shots: few | 77.7% | not dated | Language Models are Few-Shot Learners |
| 33 | text-davinci-001OpenAI | closed | Shots: 16 | 77.7% | not dated | GLaM: Efficient Scaling of Language Models with Mixture-of-Experts |
| 34 | text-davinci-001OpenAI | closed | Shots: few | 77.7% | not dated | Language Models are Few-Shot Learners |
| 35 | text-davinci-001OpenAI | closed | Shots: few | 77.7% | not dated | Using DeepSpeed and Megatron to Train Megatron-Turing NLG 530B, A Large-Scale Generative Language Model |
| 36 | mistralai/Mixtral-8x7B-v0.1Epoch: Mixtral-8x7B-v0.1, Mistral AI | open | default | 77.2% | not dated | Mixtral of Experts |
| 37 | LLaMA-65BMeta AIno page match | open | Shots: 0 | 77.0% | not dated | LLaMA: Open and Efficient Foundation Language Models |
| 38 | LLaMA-65BMeta AIno page match | open | default | 77.0% | not dated | Llama 2: Open Foundation and Fine-Tuned Chat Models |
| 39 | PaLM 62B | not stated | Shots: 0 | 77.0% | not dated | PaLM: Scaling Language Modeling with Pathways |
| 40 | PaLM 62B | not stated | Shots: 0 | 77.0% | not dated | LLaMA: Open and Efficient Foundation Language Models |
| 41 | falcon-40bTechnology Innovation Instituteno page match | open | default | 76.9% | not dated | Llama 2: Open Foundation and Fine-Tuned Chat Models |
| 42 | Qwen/Qwen2.5-Coder-14BEpoch: Qwen2.5-Coder-14B | open | default | 76.8% | not dated | Qwen2.5-Coder Technical Report |
| 43 | Llama-2-34bMeta AI | closed | default | 76.7% | not dated | Llama 2: Open Foundation and Fine-Tuned Chat Models |
| 44 | Llama-2-34bMeta AI | closed | Shots: 0 | 76.7% | not dated | Nemotron-4 15B Technical Report |
| 45 | falcon-40bTechnology Innovation Instituteno page match | open | Shots: 5 | 76.4% | not dated | Falcon2-11B Technical Report |
| 46 | falcon-40bTechnology Innovation Instituteno page match | open | Shots: 5 | 76.4% | not dated | Falcon2-11B Technical Report |
| 47 | LLaMA-33BMeta AIno page match | open | Shots: 0 | 76.0% | not dated | LLaMA: Open and Efficient Foundation Language Models |
| 48 | LLaMA-33BMeta AIno page match | open | default | 76.0% | not dated | Llama 2: Open Foundation and Fine-Tuned Chat Models |
| 49 | Meta-Llama-3-8BMeta AIno page match | open | Shots: 5 | 75.7% | not dated | The Llama 3 Herd of Models |
| 50 | mistralai/Mistral-7B-v0.1Epoch: Mistral-7B-v0.1, Mistral AI | open | Shots: 0 | 75.3% | not dated | Training Compute-Optimal Large Language Models |
| 51 | mistralai/Mistral-7B-v0.1Epoch: Mistral-7B-v0.1, Mistral AI | open | Shots: 0 | 75.3% | not dated | Nemotron-4 15B Technical Report |
| 52 | claude-3-sonnet-20240229Anthropic | closed | Shots: 5 | 75.1% | not dated | The Claude 3 Model Family: Opus, Sonnet, Haiku |
| 53 | Chinchilla (70B)DeepMind | closed | Shots: 0 | 74.9% | not dated | LLaMA: Open and Efficient Foundation Language Models |
| 54 | Chinchilla (70B)DeepMind | closed | Shots: 0 | 74.9% | not dated | Training Compute-Optimal Large Language Models |
| 55 | claude-3-haiku-20240307Anthropic | closed | Shots: 5 | 74.2% | not dated | The Claude 3 Model Family: Opus, Sonnet, Haiku |
| 56 | mistralai/Mistral-7B-v0.1Epoch: Mistral-7B-v0.1, Mistral AI | open | default | 74.2% | not dated | Mixtral of Experts |
| 57 | mistralai/Mistral-7B-v0.1Epoch: Mistral-7B-v0.1, Mistral AI | open | default | 74.2% | not dated | Gemma: Open Models Based on Gemini Research and Technology |
| 58 | Megatron-Turing NLG 530BMicrosoft,Nvidia | closed | Shots: 1 | 73.7% | not dated | Using DeepSpeed and Megatron to Train Megatron-Turing NLG 530B, A Large-Scale Generative Language Model |
| 59 | GLaM (MoE)Google | closed | Shots: 0 | 73.5% | not dated | GLaM: Efficient Scaling of Language Models with Mixture-of-Experts |
| 60 | microsoft/phi-1_5Epoch: phi-1_5, Microsoft | open | Shots: 0 | 73.4% | not dated | Textbooks Are All You Need II: phi-1.5 technical report |
| 61 | text-davinci-001OpenAI | closed | Shots: 1 | 73.2% | not dated | Language Models are Few-Shot Learners |
| 62 | text-davinci-001OpenAI | closed | Shots: 1 | 73.2% | not dated | GLaM: Efficient Scaling of Language Models with Mixture-of-Experts |
| 63 | text-davinci-001OpenAI | closed | Shots: 1 | 73.2% | not dated | Language Models are Few-Shot Learners |
| 64 | text-davinci-001OpenAI | closed | Shots: 1 | 73.2% | not dated | Using DeepSpeed and Megatron to Train Megatron-Turing NLG 530B, A Large-Scale Generative Language Model |
| 65 | Megatron-Turing NLG 530BMicrosoft,Nvidia | closed | Shots: 0 | 73.0% | not dated | Using DeepSpeed and Megatron to Train Megatron-Turing NLG 530B, A Large-Scale Generative Language Model |
| 66 | Megatron-Turing NLG 530BMicrosoft,Nvidia | closed | Shots: few | 73.0% | not dated | Using DeepSpeed and Megatron to Train Megatron-Turing NLG 530B, A Large-Scale Generative Language Model |
| 67 | GLaM (MoE)Google | closed | Shots: 1 | 73.0% | not dated | GLaM: Efficient Scaling of Language Models with Mixture-of-Experts |
| 68 | LLaMA-13BMeta AIno page match | open | Shots: 0 | 73.0% | not dated | LLaMA: Open and Efficient Foundation Language Models |
| 69 | LLaMA-13BMeta AIno page match | open | Shots: 0 | 73.0% | not dated | LLaMA: Open and Efficient Foundation Language Models |
| 70 | LLaMA-13BMeta AIno page match | open | default | 73.0% | not dated | Llama 2: Open Foundation and Fine-Tuned Chat Models |
| 71 | 01-ai/Yi-9BEpoch: Yi-9B | open | Shots: 5 | 73.0% | not dated | Yi: Open Foundation Models by 01.AI |
| 72 | deepseek-ai/DeepSeek-Coder-V2-Lite-BaseEpoch: DeepSeek-Coder-V2-Lite-Base | open | default | 72.9% | not dated | Qwen2.5-Coder Technical Report |
| 73 | Qwen2.5-Coder-7BAlibabano page match | open | default | 72.9% | not dated | Qwen2.5-Coder Technical Report |
| 74 | Llama-2-13bMeta AIno page match | open | default | 72.8% | not dated | Llama 2: Open Foundation and Fine-Tuned Chat Models |
| 75 | Llama-2-13bMeta AIno page match | open | default | 72.8% | not dated | Gemma: Open Models Based on Gemini Research and Technology |
| 76 | Llama-2-13bMeta AIno page match | open | Shots: 0 | 72.8% | not dated | Nemotron-4 15B Technical Report |
| 77 | gemma-7bGoogle DeepMindno page match | open | Shots: 0 | 72.3% | not dated | Nemotron-4 15B Technical Report |
| 78 | Yi-6B01.AIno page match | open | Shots: 5 | 71.3% | not dated | Yi: Open Foundation Models by 01.AI |
| 79 | mpt-30bMosaicMLno page match | open | default | 71.0% | not dated | Llama 2: Open Foundation and Fine-Tuned Chat Models |
| 80 | microsoft/Phi-3-mini-4k-instructEpoch: Phi-3-mini-4k-instruct, Microsoft | open | Shots: 5 | 70.8% | not dated | Phi-3 Technical Report: A Highly Capable Language Model Locally on Your Phone |
| 81 | vicuna-13b-v1.1 | not stated | Shots: 0 | 70.8% | not dated | Textbooks Are All You Need II: phi-1.5 technical report |
| 82 | Gopher (280B)DeepMind | closed | Shots: 0 | 70.2% | not dated | Using DeepSpeed and Megatron to Train Megatron-Turing NLG 530B, A Large-Scale Generative Language Model |
| 83 | text-davinci-001OpenAI | closed | Shots: 0 | 70.2% | not dated | LLaMA: Open and Efficient Foundation Language Models |
| 84 | text-davinci-001OpenAI | closed | Shots: 0 | 70.2% | not dated | Language Models are Few-Shot Learners |
| 85 | text-davinci-001OpenAI | closed | Shots: 0 | 70.2% | not dated | GLaM: Efficient Scaling of Language Models with Mixture-of-Experts |
| 86 | text-davinci-001OpenAI | closed | Shots: 0 | 70.2% | not dated | Language Models are Few-Shot Learners |
| 87 | text-davinci-001OpenAI | closed | Shots: 0 | 70.2% | not dated | Using DeepSpeed and Megatron to Train Megatron-Turing NLG 530B, A Large-Scale Generative Language Model |
| 88 | Gopher (280B)DeepMind | closed | Shots: 0 | 70.1% | not dated | LLaMA: Open and Efficient Foundation Language Models |
| 89 | Gopher (280B)DeepMind | closed | Shots: 0 | 70.1% | not dated | Training Compute-Optimal Large Language Models |
| 90 | Gopher (280B)DeepMind | closed | Shots: 0 | 70.1% | not dated | GLaM: Efficient Scaling of Language Models with Mixture-of-Experts |
| 91 | Gopher (280B)DeepMind | closed | Shots: 0 | 70.1% | not dated | Scaling Language Models: Methods, Analysis & Insights from Training Gopher |
| 92 | LLaMA-7BMeta AIno page match | open | Shots: 0 | 70.1% | not dated | LLaMA: Open and Efficient Foundation Language Models |
| 93 | LLaMA-7BMeta AIno page match | open | Shots: 0 | 70.1% | not dated | LLaMA: Open and Efficient Foundation Language Models |
| 94 | LLaMA-7BMeta AIno page match | open | default | 70.1% | not dated | Llama 2: Open Foundation and Fine-Tuned Chat Models |
| 95 | LLaMA-7BMeta AIno page match | open | Shots: 0 | 69.6% | not dated | XGen-7B Technical Report |
| 96 | Llama-2-7bMeta AIno page match | open | default | 69.2% | not dated | Llama 2: Open Foundation and Fine-Tuned Chat Models |
| 97 | Llama-2-7bMeta AIno page match | open | default | 69.2% | not dated | Gemma: Open Models Based on Gemini Research and Technology |
| 98 | Llama-2-7bMeta AIno page match | open | Shots: 0 | 69.1% | not dated | Textbooks Are All You Need II: phi-1.5 technical report |
| 99 | gpt-3.5-turbo-1106OpenAI | closed | Shots: 5 | 68.8% | not dated | Phi-3 Technical Report: A Highly Capable Language Model Locally on Your Phone |
| 100 | mpt-7bMosaicMLno page match | open | Shots: 0 | 68.6% | not dated | XGen-7B Technical Report |
| 101 | mpt-7bMosaicMLno page match | open | default | 68.3% | not dated | Llama 2: Open Foundation and Fine-Tuned Chat Models |
| 102 | mpt-7bMosaicMLno page match | open | Shots: 0 | 68.0% | not dated | Textbooks Are All You Need II: phi-1.5 technical report |
| 103 | Qwen/Qwen2.5-Coder-3BEpoch: Qwen2.5-Coder-3B | open | default | 67.4% | not dated | Qwen2.5-Coder Technical Report |
| 104 | falcon-7bTechnology Innovation Instituteno page match | open | Shots: 0 | 67.2% | not dated | XGen-7B Technical Report |
| 105 | falcon-7bTechnology Innovation Instituteno page match | open | Shots: 5 | 67.2% | not dated | Falcon2-11B Technical Report |
| 106 | falcon-7bTechnology Innovation Instituteno page match | open | Shots: 5 | 67.2% | not dated | Falcon2-11B Technical Report |
| 107 | open_llama_7b | not stated | Shots: 0 | 67.0% | not dated | XGen-7B Technical Report |
| 108 | LLaMA-7BMeta AIno page match | open | Shots: 0 | 66.9% | not dated | Textbooks Are All You Need II: phi-1.5 technical report |
| 109 | falcon-7bTechnology Innovation Instituteno page match | open | default | 66.3% | not dated | Llama 2: Open Foundation and Fine-Tuned Chat Models |
| 110 | falcon-7bTechnology Innovation Instituteno page match | open | Shots: 0 | 66.2% | not dated | Textbooks Are All You Need II: phi-1.5 technical report |
| 111 | EleutherAI/gpt-neox-20bEpoch: gpt-neox-20b, EleutherAI | open | Shots: 0 | 66.1% | not dated | XGen-7B Technical Report |
| 112 | INTELLECT-1-InstructPrime Intellect,Hugging Face,Arcee AI | not stated | Shots: 5 | 65.8% | not dated | INTELLECT-1 Technical Report |
| 113 | gemma-2bGoogle DeepMindno page match | open | default | 65.4% | not dated | Gemma: Open Models Based on Gemini Research and Technology |
| 114 | meta-llama/Meta-Llama-3-8B-InstructEpoch: Meta-Llama-3-8B-Instruct, Meta AI | open | Shots: 5 | 65.0% | not dated | Phi-3 Technical Report: A Highly Capable Language Model Locally on Your Phone |
| 115 | Salesforce/xgen-7b-8k-baseEpoch: xgen-7b-8k-base, Salesforce | open | Shots: 0 | 64.9% | not dated | XGen-7B Technical Report |
| 116 | facebook/opt-13bEpoch: opt-13b | open | Shots: 0 | 64.7% | not dated | XGen-7B Technical Report |
| 117 | gpt-j-6bEleutherAI,LAIONno page match | open | Shots: 0 | 64.5% | not dated | XGen-7B Technical Report |
| 118 | bigcode/starcoder2-15bEpoch: starcoder2-15b, Hugging Face,ServiceNow,Nvidia,BigCode | open | default | 64.3% | not dated | Qwen2.5-Coder Technical Report |
| 119 | RedPajama-INCITE-7B-Base | not stated | Shots: 0 | 63.8% | not dated | XGen-7B Technical Report |
| 120 | deepseek-ai/deepseek-coder-33b-baseEpoch: deepseek-coder-33b-base, DeepSeek,Peking University | open | default | 62.0% | not dated | Qwen2.5-Coder Technical Report |
| 121 | mistralai/Mixtral-8x7B-v0.1Epoch: Mixtral-8x7B-v0.1, Mistral AI | open | Shots: 5 | 62.0% | not dated | Phi-3 Technical Report: A Highly Capable Language Model Locally on Your Phone |
| 122 | dolly-v2-12bDatabricksno page match | open | Shots: 0 | 61.8% | not dated | XGen-7B Technical Report |
| 123 | facebook/opt-1.3bEpoch: opt-1.3b, Meta AI | open | Shots: 0 | 61.0% | not dated | Textbooks Are All You Need II: phi-1.5 technical report |
| 124 | Cerebras-GPT-13BCerebras Systemsno page match | open | Shots: 0 | 60.8% | not dated | XGen-7B Technical Report |
| 125 | Qwen2.5-Coder-1.5BAlibabano page match | open | default | 60.7% | not dated | Qwen2.5-Coder Technical Report |
| 126 | Qwen/CodeQwen1.5-7BEpoch: CodeQwen1.5-7B | open | default | 59.8% | not dated | Qwen2.5-Coder Technical Report |
| 127 | gpt2-xlOpenAIno page match | open | default | 58.3% | not dated | via Epoch AI |
| 128 | gpt2-xlOpenAIno page match | open | Shots: 0 | 58.3% | not dated | Textbooks Are All You Need II: phi-1.5 technical report |
| 129 | deepseek-ai/deepseek-coder-6.7b-baseEpoch: deepseek-coder-6.7b-base, DeepSeek,Peking University | open | default | 57.6% | not dated | Qwen2.5-Coder Technical Report |
| 130 | bigcode/starcoder2-3bEpoch: starcoder2-3b, Hugging Face,ServiceNow,Nvidia,BigCode | open | default | 57.1% | not dated | Qwen2.5-Coder Technical Report |
| 131 | bigcode/starcoder2-7bEpoch: starcoder2-7b, Hugging Face,ServiceNow,Nvidia,BigCode | open | default | 57.1% | not dated | Qwen2.5-Coder Technical Report |
| 132 | gemma-7bGoogle DeepMindno page match | open | Shots: 5 | 55.6% | not dated | Phi-3 Technical Report: A Highly Capable Language Model Locally on Your Phone |
| 133 | Qwen/Qwen2.5-Coder-0.5BEpoch: Qwen2.5-Coder-0.5B | open | default | 54.8% | not dated | Qwen2.5-Coder Technical Report |
| 134 | microsoft/phi-2Epoch: phi-2, Microsoft | open | Shots: 5 | 54.7% | not dated | Phi-3 Technical Report: A Highly Capable Language Model Locally on Your Phone |
| 135 | mistralai/Mistral-7B-v0.1Epoch: Mistral-7B-v0.1, Mistral AI | open | Shots: 5 | 54.2% | not dated | Phi-3 Technical Report: A Highly Capable Language Model Locally on Your Phone |
| 136 | deepseek-ai/deepseek-coder-1.3b-baseEpoch: deepseek-coder-1.3b-base, DeepSeek,Peking University | open | default | 53.3% | not dated | Qwen2.5-Coder Technical Report |
| 137 | stabilityai/stablelm-tuned-alpha-7bEpoch: stablelm-tuned-alpha-7b | open | Shots: 0 | 51.5% | not dated | XGen-7B Technical Report |
Open-weight rows are shaded. Dates are when the result was run or evaluated, where the data says. Data from Epoch AI (CC-BY 4.0), fetched 19 Nov 2026: Epoch AI, ‘Capabilities & benchmarking’. Published online at epoch.ai. Retrieved from ‘https://epoch.ai/benchmarks’ [online resource]. Epoch AI. Method