Alibaba releases Qwen3.8-Max, 2.4T-parameter AI for week-long autonomous tasks

Qwen3.8-Max represents a significant milestone in open-weight AI development, moving beyond single-turn interactions to sustained, multi-day autonomous work. The model's architecture builds on Qwen3.5 but scales to 2.4 trillion total parameters with substantially expanded reinforcement learning training covering multi-day workflows, nested structures, and multiple agent harnesses—a departure from the typical single-task focus. This shift in training methodology appears to be the foundation for its long-horizon capability: the team's internal benchmark index rose from 0.474 to 0.725, with performance peaking around 4,000 training environments before plateauing slightly.

The concrete case studies reveal the model's range. In competitive terms, it outperformed 458 of 526 human teams in a multimodal dialogue challenge within 24 hours and beat published research by 2.7 points on the AIME24 math benchmark. In engineering contexts, it achieved an 81 percent reduction in physical chip area through iterative gate-count optimization. The e-commerce simulation—where the model quadrupled starting capital and outperformed its predecessor by more than 2.5× and a rival (GLM 5.2) by 38 percent—demonstrates planning and negotiation across hundreds of interaction rounds. These results are self-reported and not yet independently verified, though the breadth of demonstrated capabilities suggests the model moves beyond typical benchmark inflation.