Now available: Qwen3.8-Max promises days of autonomous work

Qwen has made its Qwen3.8-Max model generally available through QwenCloud, positioning it as its most capable model so far. Qwen announces in its official blog that the model has 2.4 trillion total parameters, with 95 billion active parameters, and that open weights will follow next week.

The company says Qwen3.8-Max is built on the Qwen 3.5 architecture and targets coding, business workflows, research and tasks that take hours or days rather than minutes. The focus is not simply on answering questions, but on completing multi-step work with tools, feedback and repeated checks.

Qwen highlights several internally documented demonstrations. In one, the model reportedly spent more than 10 days building an autonomous coding system that collected requirements, created issues, wrote code, ran tests and repaired failures. The associated GitHub project accumulated 265 commits, 127 pull requests and 151 issues after about 16 days of autonomous operation, according to the company.

Another test asked the model to reproduce an AI research paper and improve its data-selection method. Qwen says the model wrote about 7,600 lines of code, ran 33 GPU training rounds and eventually improved the reported result on the AIME24 mathematics benchmark by 2.7 points over the paper’s approach.

Beyond programming

Qwen also presents applications in document review, design, financial analysis, chip development and e-commerce simulations. In a chip-design benchmark, the company says Qwen3.8-Max reduced a cryptographic accelerator design from 8,298 to 678 logic gates while maintaining automated test results.

The release also expands Qwen’s multimodal agent plans. The model can work with long documents, video and visual interfaces, Qwen says, and can use visual feedback to inspect and revise its own output. Qwen is additionally introducing Qwen-MM-Plugins, a library intended to add image, video, visual-memory and specialised production tools to existing agent systems.