Qwen3-Max ranks among the most advanced language models on the market

Alibaba has launched Qwen3-Max, its largest language model to date. It features over one trillion parameters and was trained on a dataset of 36 trillion tokens. Its architecture is based on a mixture of experts, an approach that distributes tasks among specialized subcomponents, contributing to stable and efficient training. Throughout the entire process, the learning curve remained uniform, without interruptions or the need to restart or adjust the data.

Thanks to improvements in distributed computing management, the model achieves 30% more efficiency in resource usage than its predecessor. Additionally, it can handle contexts of up to one million tokens, allowing it to process extremely long documents or interactions without performance loss.

The instructional variant, Qwen3-Max-Instruct, ranks third on LMArena's Text Arena leaderboard. On SWE-Bench Verified, a test that evaluates the ability to solve real-world programming problems extracted from public repositories, it achieves 69.6%, placing it among the most competent models globally. On Tau2-Bench, designed to measure precision in tool usage by AI agents, it scores 74.8%, surpassing systems like Claude Opus 4 and DeepSeek V3.1.

Alibaba is also developing Qwen3-Max-Thinking, a version specialized in complex reasoning. Although still in training, it has already achieved perfect results on demanding mathematical tests such as AIME 25 and HMMT, by combining code execution and advanced inference strategies. The company plans to publicly release this variant in the coming months.

Qwen3-Max-Instruct is now available on the Qwen Chat platform and through the API on Alibaba Cloud. Its compatibility with the OpenAI API format facilitates its integration into existing applications. To access it, users must register on Alibaba Cloud, activate the Model Studio service, and generate an API key. The launch reinforces Alibaba's commitment to offering scalable and open artificial intelligence infrastructure to developers and researchers.

Key points

• Alibaba launches Qwen3-Max, its largest language model with over one trillion parameters trained on 36 trillion tokens.
• Qwen3-Max-Instruct ranks third on LMArena's Text Arena leaderboard.
• The model achieves 69.6% on SWE-Bench Verified and 74.8% on Tau2-Bench, surpassing Claude Opus 4 and DeepSeek V3.1.
• The mixture of experts architecture enabled stable training without interruptions or adjustments.
• Achieves 30% more efficiency in resource usage compared to its predecessor.
• Can process contexts of up to one million tokens without performance loss.
• Qwen3-Max-Thinking, a variant under development, achieves perfect results on AIME 25 and HMMT.
• Available on Qwen Chat and via API on Alibaba Cloud, compatible with OpenAI API format.