// BENCHMARK

HellaSwag benchmark

AI model leaderboard for the HellaSwag benchmark. Compare how large language models score on HellaSwag, see the full ranking, and understand what this AI benchmark measures. Qwen2.5-7B currently leads with 89.7. Commonsense sentence-completion benchmark — pick the most plausible ending out of four for a short narrative.

Leaderboard

# Model Organization Score Variant Source
#1 Qwen2.5-7B Qwen 89.7 base-cited-olmo2 official ↗
#2 Nemotron-H 56B NVIDIA 89 base-10-shot official ↗
#3 OLMo 2 32B Allen Institute for AI 88.5 base-from-hf-readme official ↗
#4 Nemotron-H 47B NVIDIA 87.9 base-10-shot official ↗
#5 Gemma 2 9B Google DeepMind 87.3 base-cited-olmo2 official ↗
#6 GPT-4o mini OpenAI 87.1 cited-phi35moe official ↗
#7 Jamba AI21 Labs 87.1 cited-qwen2-57b official ↗
#8 Mixtral 8x7B Mistral AI 86.5 cited-qwen2-57b official ↗
#9 Yi-1.5-34B 01.AI 85.9 cited-qwen2-57b official ↗
#10 Qwen2-57B-A14B Qwen 85.2 base-from-hf-readme official ↗
#11 Qwen1.5-32B Qwen 85 cited-qwen2-57b official ↗
#12 Llama 2-13B Meta AI 83.9 base-cited-olmo2 official ↗
#13 Phi-3.5-MoE Microsoft 83.8 instruct-5-shot official ↗
#14 Marin 8B Marin 83.6 base-10-shot-from-hf-readme official ↗
#15 Mistral 7B Mistral AI 83.31 v0-1-cited-falcon-mamba official ↗
#16 Nemotron-H 8B NVIDIA 83.23 base-10-shot official ↗
#17 Llama 3-8B Meta AI 82.23 base-cited-falcon-mamba official ↗
#18 Gemma 7B Google DeepMind 82.2 cited-falcon-mamba official ↗
#19 Llama 3.1-8B Meta AI 81.9 base-10-shot-cited-marin official ↗
#20 Llama 3.1-8B Meta AI 81.6 base-cited-olmo2 official ↗
#21 Gemma 2 9B Google DeepMind 80.9 instruct-cited-phi35moe official ↗
#22 Falcon Mamba Technology Innovation Institute 80.82 base-from-hf-readme official ↗
#23 Qwen2-7B Qwen 80.7 base-from-hf-readme official ↗
#24 DCLM 7B Apple 80.43 from-hf-readme official ↗
#25 Qwen2.5-7B Qwen 80 cited-phi4-mini official ↗
#26 Llama 3.2 3B Meta AI 77.2 cited-phi4-mini official ↗
#27 Mistral NeMo Mistral AI 76.7 instruct-cited-phi35moe official ↗
#28 Ministral 8B Mistral AI 74.6 cited-phi4-mini official ↗
#29 Qwen2.5-3B Qwen 74.6 cited-phi4-mini official ↗
#30 Phi-2 Microsoft 73.1 cited-qwen2 official ↗
#31 phi-3.5-mini Microsoft 72.2 cited-phi4-mini official ↗
#32 Gemma 2B Google DeepMind 71.4 cited-qwen2 official ↗
#33 Phi-4 Mini Microsoft 69.1 instruct-5-shot official ↗
#34 Gemini 1.5 Flash (Sep 2024) Google DeepMind 67.5 cited-phi35moe official ↗
#35 Qwen2-1.5B Qwen 66.6 base-from-hf-readme official ↗
#36 Qwen2-0.5B Qwen 49.3 base-from-hf-readme official ↗

Frequently asked questions about HellaSwag

What is the HellaSwag benchmark?

Commonsense sentence-completion benchmark — pick the most plausible ending out of four for a short narrative.

How is the HellaSwag benchmark scored?

HellaSwag is scored using the accuracy metric, where a higher score is better. GenAIList aggregates reported scores from model providers and papers into a single ranked leaderboard.

Which AI model scores highest on HellaSwag?

As of the latest reported scores on GenAIList, Qwen2.5-7B achieves the highest result on HellaSwag with a score of 89.7.

Is a higher HellaSwag score better?

Yes. On HellaSwag a higher score indicates better performance, so models near the top of the leaderboard are the strongest.