Alibaba launches more efficient Qwen3-Next artificial intelligence model

According to Huobi HTX, Alibaba's Tongyi Qianwen has released the next-generation basic model architecture Qwen3-Next, and has opened the Qwen3-Next-80B-A3B series models based on this architecture. Compared with Qwen3's MoE model structure, this structure has made the following core improvements: hybrid attention mechanism, high-sparseness MoE structure, a series of stable and friendly training optimizations, and a multi-token prediction mechanism that improves inference efficiency. Based on the model structure of Qwen3-Next, Alibaba trained the Qwen3-Next-80B-A3B-Base model, which has 80 billion parameters and only 3 billion parameters are activated. This Base model achieves similar or even slightly better performance to the Qwen3-32B density model, and its training cost (GPU hours) is only less than one-tenth of that of Qwen3-32B. The inference throughput in a context above 32k is more than ten times that of Qwen3-32B, achieving the ultimate training and inference cost-effectiveness.