Qwen3 Max

Qwen3 Max is the flagship member of the Qwen3 series, built on a Mixture of Experts architecture that enables stable and efficient training at extraordinary scale. With over 1 trillion parameters and pretraining on 36 trillion tokens, the model incorporates a global-batch load balancing loss to maintain smooth, consistent pretraining curves throughout its development. It is designed to excel in complex, multi-step scenarios, delivering state-of-the-art agent programming and tool usage capabilities, and ranks among the top reasoning models on competitive leaderboards. The model prioritizes accuracy in math, coding, and science tasks, supports over 100 languages with stronger translation and commonsense reasoning, and is optimized for retrieval-augmented generation workflows.

The model builds on the Qwen3 series lineage and benefits from reinforcement learning scaling that advances capabilities across factual knowledge, reasoning, instruction following, and alignment with human preferences. Its performance on 19 established benchmarks is comparable to leading models, and advanced test-time scaling techniques further boost its reasoning output. While a separate Qwen3-Max-Thinking variant explores extended deliberative processes with tool usage, the base Qwen3 Max operates without a dedicated thinking mode and focuses on reliability and consistency in production environments. For teams working in Chinese and English, handling complex instructions, open-ended Q&A, and agent orchestration, the model provides a grounded and capable foundation.