DeepSeek releases the official V3.2 version, enhancing its reasoning and tool-calling capabilities.

On December 2nd, DeepSeek's official WeChat account announced a significant update, officially launching two new models: V3.2 and V3.2-Speciale. This upgrade marks a dual breakthrough for DeepSeek in both general reasoning and vertical domain capabilities.

The V3.2 model performs exceptionally well in multiple reasoning benchmark tests, reaching a level comparable to GPT-5. This model further strengthens its tool-calling capabilities under the "thinking mode," enabling it to more autonomously and accurately connect external tools with internal reasoning processes, thereby significantly improving the generalization performance and task adaptability of the agent, providing more reliable support for automated processing in complex scenarios.

The other model, V3.2-Speciale, specifically optimized for mathematics and programming scenarios, demonstrates top-tier performance in academic competition-level tasks. In relevant evaluations of authoritative competitions such as the International Mathematical Olympiad (IMO) and the International Olympiad in Informatics (IOI), its performance has reached the level of the top ten human contestants, showcasing powerful logical reasoning, symbolic computation, and code generation capabilities.

Currently, DeepSeek's web interface, mobile app, and API services have been synchronously upgraded, allowing users to immediately experience the comprehensive enhancements of the V3.2 model. It is worth noting that the V3.2-Speciale model will be temporarily available for trial via API until December 15th, providing developers and researchers in relevant fields with a limited-time opportunity for in-depth testing.

This dual-model release reflects DeepSeek's strategic layout of continuously advancing the frontiers of AI technology while simultaneously deepening its expertise in specialized applications, and lays a solid foundation for the iteration of more scenario-specific and specialized models in the future.