Qwen3 Max

Qwen3-Max represents the flagship release of the Qwen3 series, designed to push the boundaries of large language model performance. Built upon a MoE (Mixture of Experts) architecture, the model leverages over 1 trillion parameters trained on 36 trillion tokens, incorporating a global-batch load balancing loss to maintain stable and smooth pretraining. This architectural foundation enables state-of-the-art results across knowledge, reasoning, coding, instruction following, and multilingual understanding benchmarks. The preview version already claimed the third position on the Text Arena leaderboard, surpassing GPT-5-Chat, and the official release brought further enhancements in coding and agent capabilities, positioning the model as a versatile engine for complex enterprise and research applications.

The Qwen3-Max lineage extends through substantial reinforcement learning and advanced test-time scaling techniques that significantly boost reasoning performance. On 19 established benchmarks, the model demonstrates competitive performance comparable to leading proprietary systems. Its adaptive tool-use capabilities enable on-demand retrieval and code interpreter invocation, while the thinking variant has achieved 100% on challenging benchmarks like AIME 25 and HMMT. Optimized for retrieval-augmented generation and tool calling, the model excels at following complex instructions in Chinese and English, reduces hallucinations, and delivers higher-quality responses across open-ended Q&A, writing, and conversation tasks—supporting over 100 languages with stronger translation and commonsense reasoning.