Overview

DeepSeek-V3.2 succeeded V3.2-Exp and launched in the DeepSeek app, web and API. It combines DeepSeek Sparse Attention (DSA), which cuts the cost of long contexts, with large-scale reinforcement learning and a synthetic pipeline for agentic tool-use training. DeepSeek also published V3.2-Speciale, a higher-compute variant aimed at reasoning.

Highlights

  • DeepSeek Sparse Attention for efficient long contexts
  • New chat template with tool calling inside the reasoning process
  • Speciale variant reported at gold-medal level on IMO and IOI 2025 problems
  • MIT license

Typical uses

Reasoning, coding and agent workloads on self-hosted clusters or through low-cost APIs.

Things to consider

  • At about 685B parameters it needs multi-GPU servers.
  • Speciale targets reasoning and does not support tool use.
  • Benchmark and olympiad claims are the developer’s own.

Specifications are taken from the model card on Hugging Face and the provider’s announcement. Benchmark results depend on evaluation settings; see how AI benchmarks work.