DeepSeek launches V3.2-Exp

DeepSeek just introduced V3.2-Exp, an experimental upgrade to its V3.1-Terminus model, with a sharp focus on efficiency and cost.

‍

Tech edge

• Debuts DeepSeek Sparse Attention (DSA) for faster long-context training & inference

• Fine-grained sparse attention = lower compute, minimal quality loss

• Benchmarks: performance on par with V3.1-Terminus

‍

API update

• Available now on App, Web & API

• API prices cut by 50%+, effective immediately

‍

DeepSeek is betting on efficiency as the competitive lever, not just scaling bigger models, but making them cheaper and faster to use.