DeepSeek Unveils Two New AI Models
According to DeepSeek's official WeChat account, the company has officially released two new models today: DeepSeek-V3.2 and DeepSeek-V3.2-Speciale.
DeepSeek-V3.2 is designed to balance reasoning capabilities with output length, making it ideal for everyday use. In recent public reasoning benchmarks, V3.2 achieved performance on par with GPT-5, slightly below Gemini-3.0-Pro. Compared with Kimi-K2-Thinking, it significantly reduces output length, cutting both computational costs and user wait times—making it more efficient for routine AI tasks.
For users seeking advanced reasoning and complex problem-solving, DeepSeek-V3.2-Speciale pushes the boundaries of open-source AI. This enhanced long-thinking version of V3.2 incorporates DeepSeek-Math-V2's theorem-proving abilities. Impressively, V3.2-Speciale has already secured gold medals at major competitions, including IMO 2025, CMO 2025, ICPC World Finals 2025, and IOI 2025, demonstrating its exceptional problem-solving power.
Currently, DeepSeek-V3.2 is fully available through the official web portal, mobile app, and API. The Speciale version is accessible via a temporary API for community testing and research, offering early access to its advanced capabilities.
