DeepSeek-V3.2 is an open-weights large language model from DeepSeek AI that is positioned around efficient reasoning and agentic use cases. Its creator publishes the model on Hugging Face under an MIT-licensed repository alongside a companion Technical Report, signaling a research-driven release intended for developers who want to inspect or self-host the weights rather than rely solely on a hosted endpoint.

The defining technical contribution of DeepSeek-V3.2 is DeepSeek Sparse Attention (DSA), an attention mechanism that lowers computational cost while preserving model quality, with a design that specifically targets long-context scenarios. The release also highlights a scalable reinforcement-learning post-training pipeline, and DeepSeek AI states that with this protocol scaled, V3.2 performs comparably to GPT-5 on reasoning benchmarks. A separate sibling variant, DeepSeek-V3.2-Speciale, is published as its own repository, suggesting a tiered family where the base model emphasizes efficiency and the Speciale variant explores additional capability.