Sep 30, 2025 03:08 PM (GMT+8) · EqualOcean
On September 30, Chinese AI developer DeepSeek released its latest experimental large model, V3.2-Exp, on Hugging Face. The model outperforms its predecessor in long-text handling and training efficiency, and introduces a sparse attention mechanism that significantly lowers compute costs. The company says this version serves as a transitional product toward the next-generation architecture and announced that API prices will be cut by more than 50%. Analysts believe that if DeepSeek can maintain the cost-performance advantages of V3 and R1, it could still pose competitive pressure on domestic rivals such as Tongyi Qwen and the U.S. OpenAI.

Source: