Alibaba unveils Qwen3-Max, a model with over a trillion parameters

Alibaba Qwen/Qwen models

At its Apsara Conference, Alibaba Cloud launched Qwen3-Max, its largest language model with over a trillion parameters. Alibaba said it is strongest at writing code and agentic tasks; it was released alongside Qwen3-Omni and other specialised models.

• Trained on 36 trillion tokens; achieves 74.8 on Tau2-Bench for tool use and multi-step tasks, surpassing Claude Opus 4 and DeepSeek V3.1.
• Available immediately via Alibaba Cloud API and Qwen Chat in Instruct (live) and Thinking (in training) variants with tiered pricing.
• Alibaba announced expanded investment in AI infrastructure beyond its previous three-year RMB 380B ($53B) program.