Alibaba has introduced its latest AI model architecture, Qwen3-Next, through its Tongyi Qianwen division. The new model, Qwen3-Next-80B-A3B, is open-sourced and boasts significant improvements over the previous Qwen3 MoE architecture. Key enhancements include a hybrid attention mechanism, a highly sparse MoE structure, and a multi-token prediction mechanism, all contributing to more efficient inference. The Qwen3-Next-80B-A3B-Base model, built on this architecture, features 80 billion parameters but activates only 3 billion, achieving performance on par with or better than the Qwen3-32B dense model. Notably, its training costs are less than one-tenth of the Qwen3-32B, and its inference throughput for contexts above 32k is over ten times greater, offering substantial cost-effectiveness in both training and inference.