Arena-Hard - release-reported (judge not official or not stated)

Chat & Writing

Arena-Hard(-Auto) win-rate style scores reported by a model release that does not use, or does not state, the official Arena-Hard-Auto v0.1 setup (GPT-4-Turbo judge against the GPT-4-0314 baseline); for example Mistral's GPT-4o-judged runs. Kept as a diagnostic because judge and baseline change the scale. Higher is better.

Top models (higher is better)

Model	Score
Mistral Medium 3	97.1
GPT-4o (source unspecified)	95.4
Command A	95.1
Claude 3.7 Sonnet	93.2
Llama 4 Maverick Instruct	91.8
GPT-4o Mini	89.7
Mistral Small 3 24B Instruct 2501	87.3
Qwen2.5 Instruct 32B	86.0
Llama 3.3 70B Instruct	84.0
Gemma 2 27B IT	78.8
Ministral 8B Instruct 2410	70.9
Gemma 2 9B IT	68.7
Ministral 3B	64.3
Llama 3.1 8B Instruct	62.4
Ministral 3 14B Instruct 2512	55.1
Qwen3-VL-8B-Instruct	52.8
Gemma 2 2B IT	51.7
Ministral 3 8B Instruct 2512	50.9
Llama 3.2 3B Instruct	46.0
Mistral 7B Instruct v0.3	44.3
Qwen3 VL 4B Instruct	43.8
Gemma 3 12B IT	43.6
Qwen3 14B	42.7
Gemma 3 4B IT	31.8
Ministral 3 3B Instruct 2512	30.5