DeepSeek V3.2

DeepSeek-V3.2 is the official successor to V3.2-Exp, released in December 2025 as a reasoning-first model purpose-built for agentic workloads. It is the first model to integrate thinking directly into tool use, supporting tool calls in both thinking and non-thinking modes without swapping models mid-flow. The release introduces DeepSeek Sparse Attention (DSA), a new attention mechanism designed for efficient long-context inference across agent trajectories. Training leaned on a massive new synthesis pipeline spanning more than 1,800 environments and 85,000 complex instructions. The sibling V3.2-Speciale variant pushes reasoning further and benchmarks competitively with GPT-5 and Gemini 3.0 Pro. Weights are open on Hugging Face and the hosted API is available through DeepSeek directly and via OpenRouter, continuing DeepSeek's pattern of shipping frontier-adjacent capability at a fraction of closed-model pricing.