WildBench - release-reported score

Name: WildBench - release-reported score
Published: 2024-06-07

Chat & Writing · 2024-06-07

WildBench evaluates chat models on challenging real-user tasks with checklist-guided LLM judging. This row holds the WB-Score-style number a model release prints when it does not state the WildBench version or uses its own judge (for example GPT-4o). Higher is better.

Top models (higher is better)

Model	Score
Ministral 3 14B Instruct 2512	68.5
Ministral 3 8B Instruct 2512	66.8
Qwen3-VL-8B-Instruct	66.3
Qwen3 14B	65.1
Gemma 3 12B IT	63.2
Ministral 3 3B Instruct 2512	56.8
Qwen3 VL 4B Instruct	56.8
GPT-4o Mini	56.1
Qwen2.5 Instruct 32B	52.7
Mistral Small 3 24B Instruct 2501	52.3
Llama 3.3 70B Instruct	50.0
Gemma 3 4B IT	49.1
Gemma 2 27B IT	48.2
Gemma 2 9B IT	43.8
Mistral NeMo Instruct 2407	42.6
Ministral 8B Instruct 2410	41.3
Llama 3.1 8B Instruct	37.0
Ministral 3B	36.3
Mistral 7B Instruct v0.3	33.1
Gemma 2 2B IT	32.5
Llama 3 8B Instruct	28.8
Llama 3.2 3B Instruct	27.2