Mixture of Qwen 3 VL8B and Kimi K2.5 beats GPT and Gemini on Video Web Arena

A mixture of Qwen 3 VL8B and Kimi K2.5 achieved SOTA on Video Web Arena, outperforming GPT models by 18% and Gemini models by 25%, while costing 3.7x less and running 3x faster. The approach decomposes visual web navigation into subtasks suited to different models.