Alibaba launches more efficient Qwen3-Next artificial intelligence model

Tongyi Qianwen, a subsidiary of Alibaba, has released the next-generation basic model architecture Qwen3-Next and open-sourced the Qwen3-Next-80B-A3B series of models based on this architecture. This structure has the following core improvements compared to the MoE model structure of Qwen3: mixed attention mechanism, high sparsity MoE structure, a series of training stability-friendly optimizations, and a multi-token prediction mechanism to improve inference efficiency. Based on the model structure of Qwen3-Next, Alibaba trained Qwen3-Next-80B-A3B-Base model, which has 80 billion parameters only activate 3 billion parameters. The Base model achieves similar or even slightly better performance than the Qwen3-32B dense model, while its training cost (GPU hours) is only less than one-tenth of that of Qwen3-32B, and the inference throughput in the context of 32k or more is more than ten times that of Qwen3-32B, achieving the ultimate cost-effectiveness of training and inference.