AlignBench v1.1

Name: AlignBench v1.1
Published: 2024-06-15

Chat & Writing · 2024-06-15

AlignBench is a Chinese alignment benchmark of user queries across eight categories, each answer graded by a GPT-4-class judge on a 1-10 scale with a reference answer and a multi-dimensional rubric. This row is the 683-query v1.1 release of June 15, 2024, whose factual reference answers were checked and corrected; scores are the mean judge rating. Results that predate v1.1 or do not name it are the separate alignbench.

Top models (higher is better)

Model	Score
Qwen3-235B-A22B	8.9
o1 (2024-12-17)	8.9
DeepSeek-R1	8.8
Qwen3 32B	8.7
QwQ-32B	8.7
Qwen3-30B-A3B	8.7
DeepSeek-V3	8.6
Qwen3 14B	8.6
Qwen3 8B	8.5
GPT-4o (2024-11-20)	8.4
o3-mini	8.4
Qwen3 4B	8.3
Claude 3.5 Sonnet (June 2024)	8.2
Qwen2 72B Instruct	8.2
Qwen-Turbo (rolling alias)	8.0
Llama 4 Maverick Instruct	8.0
Qwen2.5-Turbo	7.9
GPT-4o Mini	7.8
Gemma 3 27B IT	7.8
Gemma 3 12B IT	7.8
DeepSeek-V2.5	7.7
Mistral Large 2 (Instruct 2407)	7.7
Phi-4	7.6
Qwen3 1.7B	7.6
Llama 4 Scout Instruct	7.5
DeepSeek R1 Distill Qwen 14B	7.4
DeepSeek-R1-Distill-Qwen-32B	7.3
Gemma 2 27B IT	7.2
Qwen2 7B Instruct	7.1
Gemma 2 9B IT	7.0
Qwen2-57B-A14B Instruct	7.0
DeepSeek-R1-Distill-Llama-70B	6.7
Qwen2.5 3B Instruct	6.5
DeepSeek R1 Distill Llama 8B	6.2
Qwen3 0.6B	6.1
Phi-4-mini-instruct	6.0
Llama 3.1 405B Instruct	6.0
Llama 3.1 70B Instruct	5.9
Gemma 3 1B IT	5.3
DeepSeek-R1-Distill-Qwen-1.5B	5.0