Autonomous work now spans several days with Qwen3.8-Max

With 2.4 trillion parameters, Qwen3.8-Max targets long-running, multimodal, and coding tasks. Its open weights are set to be released soon.

Available through QwenCloud, Qwen3.8-Max has 2.4 trillion parameters, with 95 billion activated during inference. The model processes both text and images, with a context window of up to one million tokens. Its open weights are expected to be released soon on Hugging Face and ModelScope. Qwen3.8-27B is set to follow the same path.

The team is placing particular emphasis on long-running coding tasks. In an experiment conducted over more than two weeks, the model built the oh-my-cli project from an empty folder while handling requests, tests, fixes, and pull requests. At the time, the public repository contained 265 commits, 127 pull requests, and 151 issues.

Other experiments involved reproducing a scientific paper, autonomously improving its method, and taking part in a multimodal classification competition. In the latter case, Qwen3.8-Max reportedly outperformed 458 of the 526 participating teams after making 45 submissions without human intervention.

Long-running tasks also extend to hardware design and business simulation. In a chip design environment, the model completed around 500 optimization rounds and reduced a circuit from 8,298 to 678 gates. In another experiment, it was tasked with running simulated online stores for one year, handling more than 2,000 interactions involving inventory, pricing, suppliers, and cash flow.

Its multimodal capabilities remain active throughout the execution process. Qwen3.8-Max can examine intermediate results, identify a visual or functional error, and adjust its plan accordingly. This approach combines coding with interface interaction across tasks involving the web, applications, documents, video, Blender, and CAD software.

The API is priced at $2 per million input tokens, $6 per million output tokens, and $0.25 for implicit caching. Three reasoning-effort levels are available. The service provides interfaces compatible with OpenAI and Anthropic protocols, with documented integrations for Claude Code, Codex, Qwen Code, Qoder, and OpenClaw.