MT-Bench - release-reported (non-standard or unstated judge)

Chat & Writing

MT-Bench scores (80 two-turn questions, 1-10 judge scale) reported by a model release that grades with a judge other than the official GPT-4 setup (for example GPT-4o as judge, or a 'dev' variant) or does not state the judge. Judge choice shifts the scale, so these values are kept as a diagnostic apart from official MT-Bench.

Top models (higher is better)

Model	Score
Mistral Small 3 24B Instruct 2501	8.3
GPT-4o Mini	8.3
Ministral 8B Instruct 2410	8.3
Qwen2.5 Instruct 32B	8.3
Ministral 3B	8.1
Llama 3.3 70B Instruct	8.0
Gemma 2 27B IT	7.9
Mistral NeMo Instruct 2407	7.8
Pixtral 12B	7.7
Gemma 2 9B IT	7.6
Gemma 2 2B IT	7.5
Llama 3.1 8B Instruct	7.5
Llama 3.2 3B Instruct	7.2
Llama 3 8B Instruct	6.8
Mistral 7B Instruct v0.3	6.7
Qwen2 VL 7B Instruct	6.4
Phi-3.5-vision-instruct	6.3