DeepSeek has unveiled DeepSeek-V3.2, a next-generation artificial intelligence model designed to balance high computational efficiency with advanced reasoning and agentic performance. The company said the model introduces multiple technical breakthroughs aimed at improving long-context understanding, reinforcement learning scalability, and tool-based reasoning.

At the core of DeepSeek-V3.2 is DeepSeek Sparse Attention (DSA), a newly developed attention mechanism that significantly reduces computational complexity while maintaining model performance. The approach is particularly optimised for long-context scenarios, enabling more efficient processing without compromising reasoning quality.

DeepSeek also introduced a scalable reinforcement learning (RL) framework, supported by increased post-training compute. According to the company, DeepSeek-V3.2 performs on par with GPT-5, while its high-compute variant, DeepSeek-V3.2-Speciale, surpasses GPT-5 and demonstrates reasoning capabilities comparable to Gemini-3.0-Pro. The model achieved gold-medal performance in the 2025 International Mathematical Olympiad (IMO) and International Olympiad in Informatics (IOI).

To strengthen agentic capabilities, DeepSeek developed a large-scale agentic task synthesis pipeline, enabling systematic generation of training data for tool-use and interactive reasoning. This improves compliance, generalisation, and real-world agent performance.

DeepSeek has also released its final submissions for IOI 2025, ICPC World Finals, IMO 2025, and CMO 2025 for community verification, alongside updates to its chat template featuring enhanced tool-calling and “thinking with tools” functionality.