DeepSeek V3.2 (Reasoning)

DeepSeek-V3.2 is a large-scale mixture-of-experts (MoE) language model designed for advanced reasoning and agentic tasks. Released in December 2025, the model succeeds the experimental V3.2-Exp and emphasizes high computational efficiency alongside frontier-level logic. It features a total of 671 billion parameters, with approximately 37 billion parameters activated per token during inference.

Technical Innovations

The model introduces DeepSeek Sparse Attention (DSA), a mechanism optimized for long-context scenarios that reduces computational complexity while maintaining performance across contexts up to 128,000 tokens. It also utilizes a Scalable Reinforcement Learning (RL) Framework and a specialized post-training compute budget to enhance performance in mathematical and logical benchmarks. These architectural choices allow the model to deliver reasoning capabilities comparable to top-tier proprietary models while maintaining a significantly lower inference cost.

Reasoning and Agency

DeepSeek-V3.2 integrates reasoning directly into tool-use through a large-scale agentic task synthesis pipeline. This enables the model to maintain a "thinking" mode while interacting with external tools, improving compliance and generalization in complex interactive environments. In competitive evaluations, the model and its high-compute variants have demonstrated gold-medal level performance in international mathematics and informatics competitions, including the IMO and IOI.