On September 12, 2025, Alibaba's Tongyi Qianwen division introduced its latest foundational model architecture, Qwen3-Next, and made the Qwen3-Next-80B-A3B series models available as open-source resources. This innovative model leverages a highly sparse Mixture of Experts (MoE) architecture, boasting a total of 80 billion parameters. Yet, during each inference process, it activates only 3 billion parameters, thereby attaining an exceptionally high level of computational efficiency.

When compared to its predecessor, Qwen3, Qwen3-Next showcases significant enhancements in several key areas. These include hybrid attention mechanisms, training optimization, and multi-token prediction mechanisms. As a result, it delivers superior performance in downstream tasks. Additionally, it reduces training costs by over 90% (less than one-tenth of the original cost) and boosts long-context reasoning throughput by more than tenfold.

It's important to note, however, that markets are inherently risky, and investments should be approached with caution. This article is not intended to serve as personalized investment advice. Users are encouraged to carefully consider their individual circumstances and take full responsibility for any investment decisions made based on the information provided herein.