DeepSeek releases DeepSeek-V3.2-Exp with sparse attention

A week after shipping the incremental V3.1-Terminus update, DeepSeek released DeepSeek-V3.2-Exp, an experimental model built on Terminus that introduced DeepSeek Sparse Attention (DSA), a mechanism for attending to tokens selectively rather than computing attention over an entire context window. DeepSeek said the technique reduced the compute cost of processing long inputs with minimal impact on output quality, and reported that V3.2-Exp performed on par with V3.1-Terminus across its benchmark suite despite the architectural change.

The efficiency gain was reflected directly in pricing: DeepSeek cut its API prices by more than 50% alongside the release, tying the new architecture’s lower inference cost to an immediate reduction in what developers paid to use the model. The “Exp” designation signalled that DeepSeek regarded the release as a trial of the sparse-attention approach on top of an already-shipped model, rather than a new flagship generation, with weights published openly as with the rest of the V3 line.

The release extended a pattern that had defined DeepSeek’s reception since the R1 launch in January 2025: demonstrating that meaningful efficiency gains in training or inference could be achieved and published openly, at a moment when much of the US industry’s public narrative rested on the premise that further capability required proportionally more compute. Sparse attention approaches of this kind had been explored in academic research before; DeepSeek’s contribution was to ship the technique in a production model and pass the resulting savings on through pricing rather than treating it as a proprietary advantage.