HellaSwag
Commonsense sentence completion. A 2019 benchmark, saturated for current models; kept for history.
As of 19 Sept 2026, GPT 4 0314 leads HellaSwag on BenchLeader with 95.3%, ahead of Llama 3.1 405B at 89.2%, across 68 model configurations with a published result.
- Published by
- Epoch AI Benchmarking Hub
- Category
- Reasoning
- Index weight
- Reference only
- Models
- 68
- 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
Commonsense sentence completion.
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.
- 1GPT 4 031495.3%
- 2Llama 3.1 405B89.2%
- 3Falcon-180B89.0%
- 4DeepSeek V388.9%
- 5Deepseek v287.1%
- 6PaLM 2-L86.8%
- 7Mixtral 8x7B86.7%
- 8davinci-00385.5%
- 9Llama 2-70B85.3%
- 10Falcon-40B85.3%
- 11Qwen2.5 72B84.8%
- 12LLaMA-65B84.2%
- 13Stable Beluga 284.1%
- 14PaLM 2-M84.0%
- 15Palm 540B83.8%
68 of 68
| # | ||||
|---|---|---|---|---|
| 1 |  GPT 4 0314OpenAI | 95.3% | 43.9 | 2023-03-14 |
| 2 |  Llama 3.1 405BMetaopen | 89.2% | 43.5 | 2024-07-23 |
| 3 |  Falcon-180BTII | 89.0% | – | 2023-09-06 |
| 4 |  DeepSeek V3DeepSeekopen | 88.9% | 44.8 | 2024-12-26 |
| 5 | Deepseek v2DeepSeekopen ↗ | 87.1% | – | 2024-05-07 |
| 6 |  PaLM 2-LGoogle | 86.8% | – | 2023-05-17 |
| 7 |  Mixtral 8x7BMistral AIopen | 86.7% | 32.1 | 2023-12-11 |
| 8 | davinci-003OpenAI | 85.5% | – | 2022-11-28 |
| 9 | Llama 2-70BMetaopen ↗ | 85.3% | 34.9 | 2023-07-18 |
| 10 | Falcon-40BTII | 85.3% | – | 2023-03-15 |
| 11 |  Qwen2.5 72BAlibabaopen | 84.8% | 42.7 | 2024-09-19 |
| 12 | LLaMA-65BMetaopen | 84.2% | – | 2023-02-24 |
| 13 |  Stable Beluga 2Stability AI | 84.1% | – | 2023-07-20 |
| 14 | PaLM 2-MGoogle | 84.0% | – | 2023-05-17 |
| 15 | Palm 540BGoogle | 83.8% | – | 2022-04-04 |
| 16 | Qwen2.5-Coder 32B InstructAlibabaopen ↗ | 83.0% | 42.6 | 2024-09-18 |
| 17 | Falcon 2 11BTII | 82.9% | – | 2024-05-09 |
| 18 | LLaMA-33BMeta | 82.8% | – | 2023-02-24 |
| 19 |  phi-3-medium 14BmediumMicrosoft | 82.4% | – | 2024-04-23 |
| 20 |  Megatron-Turing NLG 530BNVIDIA | 82.4% | – | 2021-10-11 |
| 21 | Nemotron-4 15BNVIDIA | 82.4% | – | 2024-02-26 |
| 22 | Gemma 7BGoogle | 82.2% | – | 2024-02-21 |
| 23 | PaLM 2-SGoogle | 82.0% | – | 2023-05-17 |
| 24 | davinci-002OpenAI | 81.5% | – | 2022-03-15 |
| 25 | Mistral 7B v0.3Mistral AIopen | 81.0% | 32.1 | 2023-09-27 |
| 26 | Llama 2-13BMetaopen ↗ | 80.7% | 36.5 | 2023-07-18 |
| 27 | Qwen2.5-Coder-14BAlibaba | 80.2% | – | 2024-09-18 |
| 28 | InstructGPT 175BOpenAI | 79.3% | – | 2022-01-27 |
| 29 | LLaMA-13BMeta | 79.2% | – | 2023-02-24 |
| 30 | OPT-175BMeta | 79.1% | – | 2022-05-02 |
| 31 | Falcon-7BTII | 78.1% | – | 2023-04-24 |
| 32 |  internlm-20bShanghai AI Lab | 78.1% | – | 2023-09-18 |
| 33 | Llama 2-7BMetaopen ↗ | 77.2% | 34.2 | 2023-07-18 |
| 34 | Phi 3 SmallMicrosoft | 77.0% | – | 2024-04-23 |
| 35 | Qwen2.5-Coder 7B InstructAlibabaopen ↗ | 76.8% | – | 2024-09-18 |
| 36 | Phi 3 MiniMicrosoftopen | 76.7% | 36.2 | 2024-04-23 |
| 37 |  MPT-7BMosaicML | 76.4% | – | 2023-05-05 |
| 38 |  Yi-9B01.AI | 76.4% | – | 2024-03-01 |
| 39 | LLaMA-7BMeta | 76.2% | – | 2023-02-24 |
| 40 | OPT-66BMeta | 74.5% | – | 2022-05-03 |
| 41 | Yi 6B01.AI | 74.4% | – | 2023-11-02 |
| 42 |  XGen-7BSalesforce | 74.2% | – | 2023-06-27 |
| 43 | open_llama_7bUnknown | 71.8% | – | 2023-06-07 |
| 44 |  INTELLECT-1Prime Intellect | 71.4% | – | 2024-11-29 |
| 45 | Gemma 2BGoogle | 71.4% | – | 2024-02-21 |
| 46 | Qwen2.5-Coder-3BAlibaba | 70.9% | – | 2024-09-18 |
| 47 |  Dolly v2 12BDatabricks | 70.8% | – | 2023-04-11 |
| 48 |  Baichuan2-13BBaichuan | 70.8% | – | 2023-09-06 |
| 49 | internlm-7bShanghai AI Lab | 70.6% | – | 2023-07-05 |
| 50 | GPT-NeoX-20BEleutherAI | 70.5% | – | 2022-04-07 |
| 51 | RedPajama-INCITE-7B-BaseUnknown | 70.3% | – | 2023-05-04 |
| 52 | opt-13bUnknown | 69.9% | – | 2022-05-11 |
| 53 | Baichuan 2-7BBaichuan | 68.0% | – | 2023-09-20 |
| 54 | InstructGPT 6BOpenAI | 67.6% | – | – |
| 55 | GPT-J-6BEleutherAI | 66.2% | – | 2021-08-05 |
| 56 | Qwen2 5 Coder 1 5BAlibaba | 61.8% | – | 2024-09-18 |
| 57 |  Cerebras-GPT-13BCerebras | 59.4% | – | 2023-03-20 |
| 58 |  vicuna-13b-v1.1LMSYS | 57.8% | – | 2023-04-12 |
| 59 |  chatglm2-6bZhipu AI | 57.0% | – | 2023-06-24 |
| 60 | InstructGPT 1.3BOpenAI | 56.1% | – | – |
Cite as: BenchLeader, “HellaSwag leaderboard”, https://www.benchleader.com/benchmarks/hellaswag, data as of 19 Sept 2026.
HellaSwag: questions
- What does HellaSwag measure?
- Commonsense sentence completion. Scores are reported in percent of tasks solved; higher is better.
- Which AI model leads HellaSwag?
- GPT 4 0314 leads HellaSwag with 95.3% as of 19 Sept 2026, ahead of Llama 3.1 405B at 89.2%.
- How many models have HellaSwag results?
- 68 model configurations have a HellaSwag result on BenchLeader, all taken from Epoch AI Benchmarking Hub.
- Who runs HellaSwag and how often is it updated?
- HellaSwag is published by Epoch AI Benchmarking Hub. BenchLeader re-reads the published results every morning and records the date each result was published.
- Does HellaSwag count toward the BenchLeader Index?
- No. HellaSwag is shown for reference but left out of the composite index.