// BENCHMARK

AlpacaEval 2 benchmark

AI model leaderboard for the AlpacaEval 2 benchmark. Compare how large language models score on AlpacaEval 2, see the full ranking, and understand what this AI benchmark measures. DeepSeek-V3 currently leads with 66.5%. GPT-4-judge LLM preference benchmark — length-controlled.

Leaderboard

# Model Organization Score Variant Source
#1 DeepSeek-V3 DeepSeek 66.5% official ↗
#2 GPT-4o (Nov 2024) OpenAI 65% lc-win-cited-tulu3 official ↗
#3 dots.llm1 Rednote 64.4% GPT-4o judge official ↗
#4 Granite 3.2 8B IBM 61.19% instruct-from-hf-readme official ↗
#5 EXAONE 3.5 32B LG AI Research 60.6% instruct-from-hf-readme official ↗
#6 EXAONE 3.5 7.8B LG AI Research 54.2% instruct-from-hf-readme official ↗
#7 DeepSeek-V3 DeepSeek 53.5% lc-win-cited-tulu3 official ↗
#8 Gemma 2 27B Google DeepMind 52.2% instruct-cited-exaone35 official ↗
#9 Tulu 3 405B Allen Institute for AI 51.4% instruct-lc-win official ↗
#10 DeepSeek-V2.5 DeepSeek 50.5% from-hf-readme official ↗
#11 Qwen2.5 Instruct (72B) Qwen 47.7% instruct-lc-win-cited-tulu3 official ↗
#12 Gemma 2 9B Google DeepMind 43.7% instruct-lc-win-cited-tulu3 official ↗
#13 Granite-4.0-H-Small IBM 42.48% official ↗
#14 Qwen2.5-32B Qwen 41% instruct-cited-exaone35 official ↗
#15 Llama 3.1-405B Meta AI 38.5% instruct-lc-win-cited-tulu3 official ↗
#16 EXAONE 3.5 2.4B LG AI Research 37.1% instruct-from-hf-readme official ↗
#17 Yi-1.5-34B 01.AI 34.8% instruct-cited-exaone35 official ↗
#18 Granite 3.2 2B IBM 34.51% instruct-from-hf-readme official ↗
#19 Tulu 3 8B Allen Institute for AI 34.5% instruct-lc-win official ↗
#20 Llama 3.1-70B Meta AI 33.4% instruct-lc-win-cited-tulu3 official ↗
#21 Qwen2.5-7B Qwen 31.7% instruct-cited-exaone35-7b official ↗
#22 Granite-4.0-H-Micro IBM 31.49% official ↗
#23 Ministral 8B Mistral AI 31.4% instruct-lc-win-cited-tulu3 official ↗
#24 Granite-4.0-H-Tiny IBM 30.61% official ↗
#25 Granite 3.1 8B IBM 30.34% instruct-cited-granite-3-2 official ↗
#26 Qwen2.5-7B Qwen 30.34% instruct-cited-granite-3-2 official ↗
#27 Hermes 3 405B Nous Research 30.2% lc-win-cited-tulu3 official ↗
#28 Gemma 2 2B Google DeepMind 29.1% instruct-cited-exaone35 official ↗
#29 Hermes 3 70B Nous Research 28.4% lc-win-cited-tulu3 official ↗
#30 Llama 3.1-8B Meta AI 27.22% instruct-cited-granite-3-2 official ↗
#31 Granite 3.1 2B IBM 27.17% instruct-cited-granite-3-2 official ↗
#32 Llama 3.1-8B Meta AI 25.7% instruct-cited-exaone35 official ↗
#33 Llama 3.1-8B Meta AI 24.2% instruct-lc-win-cited-tulu3 official ↗
#34 DeepSeek-R1-Distill-Llama-8B DeepSeek 21.85% cited-granite-3-2 official ↗
#35 Llama 3.2 3B Meta AI 18.7% instruct-cited-exaone35 official ↗
#36 Qwen2.5-3B Qwen 17.4% instruct-cited-exaone35 official ↗
#37 DeepSeek-R1-Distill-Qwen-7B Qwen 15.35% cited-granite-3-2 official ↗
#38 DeepSeek-R1-Distill-Qwen-7B DeepSeek 15.35% cited-granite-3-2 official ↗
#39 Qwen2.5-1.5B Qwen 8.4% instruct-cited-exaone35 official ↗

Frequently asked questions about AlpacaEval 2

What is the AlpacaEval 2 benchmark?

GPT-4-judge LLM preference benchmark — length-controlled.

How is the AlpacaEval 2 benchmark scored?

AlpacaEval 2 is scored using the win-rate (%) metric, where a higher score is better. The maximum achievable score is 100.000. GenAIList aggregates reported scores from model providers and papers into a single ranked leaderboard.

Which AI model scores highest on AlpacaEval 2?

As of the latest reported scores on GenAIList, DeepSeek-V3 achieves the highest result on AlpacaEval 2 with a score of 66.5%.

Is a higher AlpacaEval 2 score better?

Yes. On AlpacaEval 2 a higher score indicates better performance, so models near the top of the leaderboard are the strongest.