DeepSeek-V3.2

DeepSeek-V3.2 is a model that harmonizes high computational efficiency with superior reasoning and agent performance. DeepSeek's approach is built upon three key technical breakthroughs: DeepSeek Sparse Attention (DSA), scalable reinforcement learning framework, and large scale agentic task synthesis pipeline.

View model card in Model Garden

Model ID	deepseek-v3.2-maas
Launch stage	GA
Supported inputs & outputs	Inputs: Text, Documents Outputs: Text
Capabilities	Supported Batch predictions Function calling Structured output Not supported
Usage types	Supported Dynamic shared quota Provisioned Throughput Not supported Fixed quota
Versions	deepseek-v3.2-maas Launch stage: GA Release date: December 10, 2025
Supported regions
	Model availability	Global global
	ML processing	United States Multi-region
Limits	global: Max output: 65,536 Context length: 163,840
Pricing	See Pricing.