DeepSeek V3.2

DeepSeek V3.2 – Sparse Attention MoE with Reasoning-First Agentic Design

DeepSeek V3.2 is DeepSeek's flagship open-weight model built on three documented technical breakthroughs: DeepSeek Sparse Attention (DSA) for long-context efficiency, a scalable reinforcement learning framework for post-training compute scaling, and a large-scale agentic task synthesis pipeline that integrates reasoning directly into tool-use - the first open-source model to do so.

Architecture

DeepSeek V3.2 is a 671B total parameter, 37B active parameter Mixture-of-Experts model. The MoE layer uses 256 expert networks per layer (up from 160 in V2), activating 8 per token: 1–2 shared experts handling common patterns plus 6–7 routed experts. Built on the same model structure as DeepSeek-V3.2-Exp.

Reasoning Workflow

DeepSeek V3.2 is documented as DeepSeek's first model to integrate thinking directly into tool-use, supporting two parallel modes:

• Thinking mode: Long-chain reasoning prior to and during tool calls
• Non-thinking mode: Fast responses without chain-of-thought

A new developer role is introduced in the chat template, dedicated exclusively to search agent scenarios