DeepSeek-V3.2 introduces Sparse Attention and agentic synthesis, surpassing GPT-5

DeepSeek has released DeepSeek-V3.2, an open large language model that combines high computational efficiency with advanced reasoning capabilities. The release highlights three main technical advancements: a new sparse attention mechanism, a scalable reinforcement learning framework, and a pipeline for synthesizing agentic training data.

• DeepSeek Sparse Attention (DSA) reduces computational complexity while maintaining performance in long-context scenarios.
• A robust reinforcement learning protocol allows the model to perform comparably to GPT-5.
• The high-compute variant, DeepSeek-V3.2-Speciale, surpasses GPT-5 and matches Gemini-3.0-Pro in reasoning.
• DeepSeek-V3.2-Speciale achieved gold-medal performance in both the 2025 International Mathematical Olympiad (IMO) and the International Olympiad in Informatics (IOI).
• A large-scale agentic task synthesis pipeline improves generalization and instruction-following in complex interactive environments.

These improvements facilitate scalable agentic post-training, yielding substantial gains in robustness within complex, interactive environments.