DeepSeek’s latest experimental AI model, called DeepSeek V3.2 Exp, is here. It is built on the company’s previous V3.1-Terminus capabilities. The new model introduces DeepSeek Sparse Attention (DSA), which is designed to speed up training and inference, particularly for long context tasks without compromising on output quality.
DSA offers fine-grained sparse attention, which reduces compute costs and enhances performance when handling large amounts of data. Early benchmarks indicate that V3.2-Exp performs on par with V3.1-Terminus, making it a promising upgrade for developers and businesses relying on long-context AI tasks.
