Alibaba’s AI coded for 16 days straight and every commit is on GitHub

Alibaba has unveiled Qwen3.8-Max, a powerful multimodal model with 2.4 trillion parameters, designed to tackle complex tasks that may span several days to complete. The model, now accessible via QwenCloud and Alibaba Cloud Model Studio, is priced at $2 per million input tokens and $6 per million output tokens. Qwen3.8-Max builds upon the Qwen3.5 architecture, employing a sparse mixture-of-experts design with hybrid attention.

While the model houses 2.4 trillion parameters in total, it activates around 95 billion for each token, utilizing only the necessary parts of the model for each specific task. Alibaba plans to publish the model weights on Hugging Face and ModelScope next week, marking the first Qwen-Max model with downloadable weights.

Despite the model's impressive capabilities, self-hosting remains a challenge due to the need to store and distribute the full weights across multiple high-memory GPU nodes. Consequently, Qwen3.8-Max is more suitable for large organizations and inference providers equipped with the necessary infrastructure. Alibaba's innovative approach brings the model closer to Moonshot AI's Kimi K3, which boasts 2.8 trillion parameters.

Moonshot AI is partly owned by Alibaba, creating a competitive dynamic. The launch of Qwen3.8-Max enters the same race as Kimi K3, as both models aspire to dominate the AI landscape.

During testing, Qwen3.8-Max demonstrated its prowess by building a command-line app in 16 days, reproducing a research paper's results in five days, and enhancing a chip design after approximately 500 iterations. However, these results have yet to be independently verified. The most ambitious software demo involved Qwen3.8-Max spending 16 days building "oh-my-cli," making 265 commits and opening 127 pull requests on GitHub while addressing 151 issues.

The repository, accessible under qwen-code-dev-bot/oh-my-cli, allows developers to review the code and audit trail.

The distinction between the model and its infrastructure, or "harness," is crucial for long-running jobs. The harness manages various aspects, determining whether an agent persists for 16 days or fails within 16 minutes. Alibaba trained Qwen3.8-Max on complex projects akin to those an agent might encounter, achieving similar performance when tested against QwenWork, Claude Code, Codex, OpenClaw, and Hermes.

Alibaba has provided developers with guides to connect Qwen3.8-Max to these tools, making it easier to integrate the model into existing coding environments.

Alibaba's context window limits when using Qwen3.8-Max were tested by providing the model with a research paper without any code. The model wrote about 7,600 lines of code, initiated 33 GPU training jobs, and validated six primary results from the paper. Additionally, it generated 18 more ideas, improving the reported AIME24 result by 2.7 points.

In another test, Qwen3.8-Max optimized a cryptographic circuit over 500 iterations, reducing the design from 8,298 logic gates to 678. Post-optimization, the physical area decreased by 81% when run through OpenROAD, an open-source chip design tool. These tests highlight the model's adaptability and efficiency in handling intricate tasks, although developers will need to assess its ability to retain earlier decisions and instructions as projects expand.