DeepSeek V3.2 (DeepInfra)

DeepSeek V3.2 is an open-weight large language model from DeepSeek-AI that aims to balance computational efficiency with strong reasoning and agent performance. Its central architectural innovation is DeepSeek Sparse Attention (DSA), an attention mechanism designed to substantially cut computational complexity while keeping model quality intact in long-context settings, making the model a practical choice for workloads where context length and inference cost both matter. Alongside DSA, the model is trained with a scalable reinforcement learning framework that pushes post-training compute further, with the authors reporting that the base DeepSeek-V3.2 performs comparably to GPT-5 on their evaluations.

A high-compute sibling variant, DeepSeek-V3.2-Speciale, is described in the paper as surpassing GPT-5 and matching Gemini-3.0-Pro on reasoning benchmarks, reportedly reaching gold-medal-level performance on the 2025 International Mathematical Olympiad and the International Olympiad in Informatics, though those claims come from the authors themselves. DeepSeek-V3.2 is also paired with a large-scale agentic task synthesis pipeline that systematically generates training data to teach tool-use alongside reasoning, supporting more reliable behavior in interactive, multi-step environments. The combination of open weights, efficient long-context attention, and an explicit agent training focus makes it well suited to developers building reasoning-heavy assistants and tool-using agents who want a self-hostable foundation rather than a closed API.