ArchitectureReinforcement LearningAgentsEvaluation
DeepSeek-V3.2: Pushing the Frontier of Open Large Language Models
DeepSeek AI·December 2, 2025
DeepSeek-AI
View on arXivTL;DR
Introduces DeepSeek Sparse Attention (DSA) — a lightning-indexer plus fine-grained token selection on top of MLA — that cuts long-context cost while preserving quality, plus a scaled RL framework; the “Speciale” variant reaches IMO/IOI gold level.
Why it matters
Brought trainable fine-grained sparse attention to a frontier open model — a major open-weights efficiency and reasoning milestone.