Qwen has officially released Qwen3.8-Max, its most capable model to date and its first at Max scale slated for open weights. The model is available now through QwenCloud, while its weights are due on Hugging Face and ModelScope next week. Built on the Qwen3.5 architecture, it has 2.4 trillion parameters with 95 billion active and is aimed at coding, research, professional work, and multimodal tasks.
📢Meet Qwen3.8-Max — our most capable model to date.
— Qwen (@Alibaba_Qwen) August 3, 2026
Next week, the open weights of Qwen3.8-Max will be released, and Qwen3.8-27B is also going open-weights to meet you all!🎉
Qwen3.8-Max, a new bar for coding and cowork at 2.4T parameters:
- Autonomous coding: 10+ days of… pic.twitter.com/e3YFj2hqcT
Autonomous coding sits at the center of the launch. In one company-run test, Qwen3.8-Max spent about 16 days building and maintaining the oh-my-cli project, producing 265 commits, 127 pull requests, and 151 issues through a loop of task intake, implementation, testing, and repair. In another, it recreated a research pipeline, completed 33 GPU training rounds, and developed a method that Qwen says scored 2.7 points above the paper's approach on AIME24.
Qwen is also pitching the model for long-running professional workflows. During an autonomous hardware-design test, it reduced a cryptographic accelerator from 8,298 to 678 gates over roughly 500 turns and reached timing closure at 500 MHz after physical layout. In a year-long ecommerce simulation spanning more than 2,000 rounds, it finished with ¥416,252, a 4.16-times return and 38% more than the runner-up.
Cowork: Any role, build beyond pic.twitter.com/Hggkj3DFSs
— Qwen (@Alibaba_Qwen) August 3, 2026
The multimodal system is designed to process documents longer than 200 pages and videos exceeding 100 hours, organize their contents into traceable structures, and visually inspect its own output for corrections. Qwen is also introducing Qwen-MM-Plugins, a harness extension for multimodal memory and visual tool use in areas including video editing, Blender, and CAD. RecreationBench combines code and GUI operation to rebuild black-box applications across desktop, mobile, and web platforms.
The Qwen team is pairing the release with familiar deployment routes for developers, researchers, and enterprises. QwenCloud supports OpenAI-compatible chat completions and Responses APIs alongside an Anthropic-compatible interface, allowing use with popular agent frameworks and coding assistants. API users can select xhigh, medium, or low reasoning effort, with thinking preserved by default. Releasing the weights next week will bring a Max-scale Qwen model into self-hosted and research settings for the first time.
