Alibaba’s newest AI model operated for 16 days without human input, generating 265 commits and 127 pull requests on GitHub before completing a functional command-line utility.
The 16-day coding marathon
The model, Qwen3.8-Max, built “oh-my-cli,” a tool now live in a public repository under the account qwen-code-dev-bot. Every commit, issue, and pull request appears in the audit trail, letting developers inspect the process without depending on Alibaba’s statements alone.
Over the 16 days, the system addressed 151 GitHub issues. Alibaba has not shared enough data for outsiders to reproduce the results, though the public repository provides an unusual glimpse into how an AI agent manages a long-term software project.
Additional experiments tested the model further. Qwen3.8-Max spent five days reproducing a research paper’s results and improved a chip design after about 500 iterations.
Related: AI frees capital but firms lack next steps
A model designed for extended, complex work
Qwen3.8-Max is a multimodal system with 2.4 trillion total parameters, activating roughly 95 billion for each token to control computational costs while still tackling multi-day tasks. Its sparse mixture-of-experts architecture selects only the necessary parts of the model rather than running all parameters simultaneously.
The full model weights demand substantial infrastructure. Alibaba will release them on Hugging Face and ModelScope next week, marking the first Qwen-Max series model available for download. Even then, self-hosting will remain impractical for most developers, as the model requires multiple high-memory GPU nodes.
Pricing is $2 per million input tokens and $6 per million output tokens, targeting large organizations and inference providers. The model is already accessible through QwenCloud and Alibaba Cloud Model Studio, which supports OpenAI and Anthropic API formats for easy integration into existing workflows.
Alibaba’s benchmarks position Qwen3.8-Max near models from other leading providers. The company has not released enough information for outside researchers to verify these results. In Alibaba’s tests, the model performed similarly when tested with QwenWork, Claude Code, Codex, OpenClaw and Hermes.
Related: Arrow CEO: Search Won’t Be Rushed
Real-world performance is still uncertain. The model completed the 16-day coding task, but its ability to maintain long-term memory and adapt as projects expand remains unclear. A million-token context window has limits, and developers will need to observe how well the system retains earlier decisions in larger projects.
Previous models of similar scale have encountered practical constraints. Moonshot AI’s Kimi K3, with 2.8 trillion parameters, topped coding leaderboards after launch but quickly overwhelmed its GPU capacity, forcing the company to halt new subscriptions. Alibaba owns a 36% stake in Moonshot AI, so the two models are not entirely separate. Kimi K3’s open-weight release also faced GPU shortages within two days, illustrating the difficulties of deploying such large systems.
The difference between the model and its harness is key for long-running jobs. The AI decides what to do next, but the surrounding system—managing memory, errors, and resources—determines whether a project runs for 16 days or fails in 16 minutes. Alibaba trained Qwen3.8-Max on complex coding scenarios and published guides for integrating it with tools like QwenWork and Claude Code.
For now, the GitHub repository offers the clearest evidence of the model’s capabilities. Developers can review the commits, pull requests, and issues to see how the AI solved problems, adjusted its approach, and delivered a working tool. Whether this translates to reliable, long-term software maintenance will only become clear once the model is tested beyond Alibaba’s controlled environment.
