DeepSeek-V3.2

An efficient Mixture-of-Experts language model with 671B total parameters, featuring DeepSeek Sparse Attention (DSA) for enhanced reasoning, agentic performance, and long-context efficiency, comparable to frontier models like GPT-5.