On September 24, 2025, the 2025 Cloud Town Conference kicked off with the much-anticipated official launch of Ali Tongyi's premier model, Qwen3-Max. This cutting-edge model not only outperforms GPT-5, Claude Opus 4, and other leading counterparts but also secures a spot among the top three globally. Qwen3-Max comes in two significant versions: Instruct and Thinking. The preview iteration has already claimed the third position on the Chatbot Arena leaderboard, with expectations that the official version will elevate performance even further.

Qwen3-Max is built on a massive pre-training dataset comprising 36 trillion tokens and boasts over a trillion total parameters. It showcases robust programming skills and exceptional capabilities in invoking Agent tools. In the SWE-Bench Verified test, the Instruct version soared into the global first tier, scoring an impressive 69.6. Meanwhile, in the Tau2-Bench test, it achieved a remarkable 74.8, outperforming both Claude Opus4 and DeepSeek-V3.1.

Its reasoning-augmented version, Qwen3-Max-Thinking-Heavy, has made history by scoring a flawless 100 in the mathematical reasoning tests AIME 25 and HMMT, marking the first time such a feat has been accomplished domestically.

At present, users have the opportunity to experience Qwen3-Max free of charge on Tongyi Qianwen's QwenChat or can access its API services through Alibaba Cloud's BaiLian Platform.