Atlas

Benchmarks

← All benchmarks

MT-Bench

Chat & Writing · 2023-06-09

MT-Bench (LMSYS) is a set of 80 two-turn open-ended questions across writing, roleplay, reasoning, math, coding, extraction, STEM and humanities. The official protocol has GPT-4 grade each answer on a 1-10 scale and reports the mean score over both turns. Higher is better.

Top models (higher is better)

ModelScore
Qwen2.5 72B Instruct9.3
Qwen Plus (rolling alias)9.3
Qwen2.5 Instruct 32B9.2
Qwen2 72B Instruct9.1
Gemma 2 27B IT9.1
Llama 3.1 405B Instruct9.1
DeepSeek-V2.59.0
DeepSeek-V2-Chat9.0
Qwen-Max-04289.0
DeepSeek-Coder-V2-Instruct-07248.9
Qwen2.5 14B Instruct 1M8.9
Qwen1.5 110B Chat8.9
Qwen2.5 14B Instruct8.9
Qwen-Turbo (rolling alias)8.9
DeepSeek-V2-Chat-06288.8
Qwen2.5-Turbo8.8
Llama 3.1 70B Instruct8.8
DeepSeek-Coder-V2-Instruct8.8
Qwen2.5 7B Instruct8.8
Mixtral 8x22B Instruct8.7
Qwen1.5 72B Chat8.6
Mistral Medium 1.08.6
Mistral Large 2 (Instruct 2407)8.6
Qwen2-57B-A14B Instruct8.6
Yi 1.5 34B Chat8.5
Gemma 2 9B IT8.5
DeepSeek LLM 67B Chat8.3
GPT-3.5 Turbo 11068.3
Mixtral 8x7B Instruct8.3
Qwen1.5 32B Chat8.3
Qwen2.5 7B Instruct 1M8.3
Llama 3.1 8B Instruct8.2
Yi 1.5 9B Chat8.2
Llama 3 8B Instruct8.1
Qwen1.5 14B Chat7.9
DeepSeek-Coder-V2-Lite-Instruct7.8
Mistral 7B Instruct v0.27.6
Qwen1.5 7B Chat7.6
DeepSeek-V2-Lite-Chat7.4
Qwen1.5 MoE A2.7B Chat7.2
Loading Atlas data…