DeepSeek-AI: V2 multi-head latent attention and efficient MoE model

What changed

Compressed key-value state into a joint latent representation for efficient decoding and integrated it into the DeepSeek-V2 expert model.

Scientific scope

Compressed key-value state into a joint latent representation for efficient decoding and integrated it into the DeepSeek-V2 expert model. DeepSeekMoE and GRPO have separate earlier sources. Cache and throughput comparisons use the paper's models and settings; no universal percentage saving is claimed.

Evidence

The linked original sources describe the stated contribution and its evidence limits. WRS has not independently rerun the experiments or certified a proof. DeepSeekMoE and GRPO have separate earlier sources. Cache and throughput comparisons use the paper's models and settings; no universal percentage saving is claimed.

Contributors

• Source-listed authors and technical contributors · SCIENCE: Source-listed authors or explicitly credited technical contributors for the cited versions. Family entries retain source-specific attribution; a combined list is not one original-paper byline or equal individual work. Names alone do not establish resolved identities.

• DeepSeek-V2: A Strong, Economical, and Efficient Mixture-of-Experts Language Model
• DeepSeek-AI: V2 multi-head latent attention and efficient MoE model — official research record

Provisional WRS assessment, edition 2026-10-16.1. Independent expert calibration is pending. This grade does not certify correctness or assign ownership.

Full scientific record · Browse contribution grades · Request a correction