Changjiang Securities·2025-09-14 14:38Investment Rating - The report maintains a "Positive" investment rating for the industry [7] Core Insights - On September 12, Alibaba released the next-generation foundational model architecture Qwen3-Next and open-sourced the Qwen3-Next-80B-A3B series models, showcasing significant breakthroughs in model architecture that demonstrate the continuous advancement of domestic large models towards world-leading levels [2][4] - The performance improvements of Qwen3-Next come with a substantial reduction in training costs, which is expected to accelerate the deployment of domestic AI applications and drive a surge in computing power demand [2][4] Summary by Sections Event Description - The report details the release of Qwen3-Next and its open-sourced model series on September 12, highlighting the advancements in model architecture [4] Model Innovations - Qwen3-Next features several innovations, including: 1. A mixed attention mechanism that combines Gated DeltaNet and standard attention to balance performance and efficiency 2. A high sparsity MoE structure with a total parameter count of 80 billion, activating only about 3 billion parameters during inference 3. Stability optimizations to prevent weight growth and ensure numerical stability 4. A multi-token prediction mechanism that enhances overall model performance [9] Recommendations - The report suggests focusing on: 1. The Chinese inference computing power industry chain, particularly recommending domestic AI chip leader Cambricon 2. Alibaba Cloud's industry chain 3. Cloud service providers 4. IDC, with a focus on collaborations among major companies like Tencent, Alibaba, and ByteDance [2][9]