한국어EN
Model Architecture & Training MethodsDeepSeek-AI (Yuan et al.)DeepSeek

NSA — Native Sparse Attention

NSA — Native Sparse Attention

Why it matters

Moved sparsity from the inference stage to the pretraining stage

Source — Primary sources first; secondary reports are labelled “Press”

Paperarxiv.orghttps://arxiv.org/abs/2502.11089

Added in Week 35, 2026 · See that week

Same week

Korea temporarily suspends new DeepSeek downloadsPressNie et al. present LLaDAPaperMacron announces €109 billion for French AI and €200 billion for EuropeOfficialCourt rejects fair use for AI training in Thomson Reuters v. Ross IntelligencePress

Nearby on the same axis

Nie et al. present LLaDAPaperNaver releases a new HyperCLOVA X modelOfficialDeepSeek-V3 Technical ReportPaperByte Latent TransformerPaper