Qwen2.5-Max: Exploring the Intelligence of Large-Scale Moe Model