Qwen3-Max

Qwen3-Max represents Alibaba's flagship large language model built on a Mixture of Experts architecture that distributes computation across specialized pathways, enabling efficient handling of over a trillion total parameters. The model's design incorporates a global-batch load balancing loss that stabilizes training dynamics, resulting in consistently smooth pretraining loss curves throughout the process. This architectural foundation supports state-of-the-art performance across knowledge tasks, reasoning challenges, instruction following, human preference alignment, and multilingual understanding. The model handles million-token inputs and was specifically engineered to compete with leading frontier systems in capability and benchmark standing.

The model underwent pretraining on 36 trillion tokens, an exceptionally large corpus that underpins its broad competency range. Its preview version achieved third place on the Text Arena leaderboard, surpassing GPT-5-Chat, while its coding capabilities reached a SWE-Bench Verified score of 69.6. A dedicated Thinking variant extends the base model with test-time compute scaling that iteratively refines reasoning by drawing on prior inference steps, enabling it to autonomously invoke tools like Search, Memory, and Code Interpreter during conversation. This reasoning model achieved perfect scores on mathematical benchmarks and outperformed competing frontier systems on the Humanity's Last Exam by a significant margin. The combination of massive pretraining scale, adaptive reasoning modes, and tool autonomy positions Qwen3-Max for complex agentic workflows and demanding professional applications.