Skip to content

DeepSeek Unveils V3.2-Exp: Slashing Long-Context Costs by 6x

DeepSeek's latest update revolutionizes long-context efficiency. With open-source licensing and API price cuts, it's set to shake up the AI landscape.

Here in this picture we can see a person present and we can see he is focusing his eyes on...
Here in this picture we can see a person present and we can see he is focusing his eyes on something.

DeepSeek Unveils V3.2-Exp: Slashing Long-Context Costs by 6x

Hangzhou-based AI company DeepSeek has unveiled DeepSeek-V3.2-Exp, an update that introduces DeepSeek Sparse Attention (DSA) for improved long-context efficiency. The company, owned by High-Flyer, has also slashed API prices by 50% following the release.

DeepSeek-V3.2-Exp retains the MoE and MLA stack from previous versions, but introduces a two-stage attention path. This path consists of a lightweight 'indexer' and sparse attention over a selected subset, significantly reducing decode costs. As a result, decoding at 128k is now approximately six times cheaper.

The indexer is trained to mimic the dense model's attention distribution using KL-divergence in a two-stage process. This innovation allows DeepSeek V3.2-Exp to maintain benchmark parity while improving long-context economics. The update is a drop-in replacement for RAG and long-document pipelines where quadratic attention dominates costs.

DeepSeek has also made DeepSeek-V3.2-Exp truly open source, licensing it under MIT. The company has reduced API prices by 50% consistent with the efficiency gains of DSA.

DeepSeek's latest release, DeepSeek-V3.2-Exp, brings significant improvements in long-context efficiency and cost reduction. With its open-source licensing and substantial API price cuts, the update is poised to have a notable impact on the AI landscape.

Read also:

Latest