DownloadDeepSeek-V3.2 is a cutting-edge large language model developed by DeepSeek-AI, focused on achieving high reasoning accuracy and computational efficiency for agentic tasks. It introduces DeepSeek Sparse Attention (DSA), a new attention mechanism that dramatically reduces computational overhead while maintaining strong long-context performance. Built with a scalable reinforcement learning framework, it reaches near-GPT-5 levels of reasoning and outperforms comparable models like DeepSeek-V3.1 and Gemini-3.0-Pro in advanced benchmarks. The model was notably used in competitive AI challenges such as the 2025 International Mathematical Olympiad (IMO) and IOI, achieving top-tier results. DeepSeek-V3.2 also features a large-scale agentic task synthesis pipeline, which generates training data to enhance tool-use intelligence and multi-step reasoning. It introduces a new “thinking with tools” chat template, allowing it to reason and decide when to invoke specific tools during problem solving.
Features
- Incorporates DeepSeek Sparse Attention (DSA) for efficient long-context inference
- 685B-parameter architecture optimized for reasoning and agentic tasks
- Scalable RL framework achieving GPT-5-level reasoning accuracy
- Demonstrated gold-medal results in IMO 2025 and IOI 2025 benchmarks
- Supports “thinking with tools” paradigm for structured reasoning workflows
- Introduces a developer role for advanced agent integration scenarios
- Compatible with vLLM, SGLang, and other major inference frameworks
- Fully open-source under MIT License with available training and evaluation scripts
Project Samples
