Alibaba released Qwen3.8-Max, scaling its flagship model to 2.4 trillion parameters with 95 billion active parameters. The company will open-source the weights next week, marking the first time a Qwen-Max-class model receives open weights.

The model demonstrated autonomous coding capabilities across multi-day projects without human intervention. In one test, Qwen3.8-Max built a self-evolving software harness called "oh-my-cli" over 16 days, accumulating 265 commits, 127 pull requests, and 151 issues through continuous feedback loops.

Autonomous Research and Development

Qwen3.8-Max reproduced a research paper on data selection for language model training, then improved upon the original method. Working independently for five days, the model wrote 7,600 lines of code, executed 1,100 actions, and ran 33 GPU training rounds.

The model first spent 37 hours rebuilding the paper's pipeline from scratch and confirmed its six main findings. It then evolved 18 improvement ideas across four rounds, ultimately developing a method that beats the original approach by 2.7 points on the AIME24 mathematics benchmark.

The autonomous coding harness combines issue state machines, dispatchers, and monitoring into one execution loop. After requirements enter GitHub Issues, agents claim tasks and move through ready-to-active states, triggering end-to-end tests and continuous integration checks before merging pull requests.

Technical Architecture

Built on the Qwen 3.5 architecture, the model delivers improvements across coding, research, and long-horizon tasks. The system integrates community feedback and self-test results into engineering workflows, automatically converting user requirements into executable work items.

Qwen3.8-Max is available through QwenCloud's API platform. The model's performance spans multiple benchmarks, with particular strength in mathematical reasoning and code generation tasks that require sustained attention over extended periods.

The open weights release next week will make the model accessible to researchers and developers building on the Qwen architecture.