V3.2 family introducing DeepSeek Sparse Attention for long-context efficiency. V3.2-Speciale achieves gold-medal performance on IMO and IOI 2025, surpassing GPT-5.

Outputs 2

DeepSeek-V3.2

model
Architecture MOE
Parameters 685B
Active params 37B

Variants

Name Parameters Notes
DeepSeek-V3.2-Exp Released 2025-09-29
DeepSeek-V3.2
DeepSeek-V3.2-Speciale Max reasoning variant

DeepSeek-V3.2: Pushing the Frontier of Open Large Language Models

paper

Technical report introducing DeepSeek Sparse Attention, scalable RL framework, and large-scale agentic task synthesis. 685B-parameter MoE model.

arXiv: 2512.02556

moefrontieropen-weight