DeepSeek's 'Inconspicuous Phrase' May Transform the Entire GPU Programming Ecosystem
1 week ago / Read about 0 minute
Author:小编   

On September 29, 2025, DeepSeek released the V3.2-Exp experimental model, which incorporates the DeepSeek Sparse Attention mechanism on the foundation of V3.1-Terminus to optimize long-text training and inference efficiency. Meanwhile, the official App, web version, and mini-program were updated simultaneously, with API call costs reduced by over 50%. Additionally, TileLang, the programming language mentioned in the announcement, is a new language designed specifically for AI and high-performance computing. It is driving the construction of a domestic computing power ecosystem, bridging the gap between upper-level AI applications and underlying domestic hardware.

  • C114 Communication Network
  • Communication Home