Alibaba prepares to release open weights for 2.4 trillion parameter model

Alibaba has announced the upcoming release of open weights for Qwen 3.8 Max, a flagship model featuring 2.4 trillion parameters. The move follows a period of management transition at the Qwen laboratory and positions the organization at the frontier of high-parameter open models, competing directly with the industry’s most capable closed-source systems.

The model's architecture is built for long-horizon agentic tasks, specifically in engineering and operational domains. In internal testing, the system demonstrated the ability to conduct 10 days of unattended autonomous coding and executed a complete silicon design flow for a cryptographic accelerator. This hardware design process included RTL editing, simulation, and physical layout, eventually reducing the gate count of the target chip from 8,298 to 678 while maintaining a 500 MHz clock speed.

Technically, the model utilizes a Mixture-of-Experts (MoE) architecture with approximately 95 billion active parameters per token, suggesting an activation ratio of roughly 4 percent. It supports a one-million-token context window and includes native visual feedback loops, allowing the agent to observe and interact with graphical user interfaces during execution. While currently available via API at $2.00 per million input tokens, the release of weights scheduled for next week will allow for private deployment on high-density infrastructure.