The Powerful DeepSeek-V3.2 AI Models Were Introduced
Credit: DeepSeek
Chinese startup DeepSeek has released two new open-source models with powerful reasoning capabilities—DeepSeek-V3.2 and the enhanced DeepSeek-V3.2-Speciale. The company thus confirms that China is on equal footing with American leaders OpenAI and Google.
DeepSeek-V3.2 is a stable version of V3.2-Exp, released in September. The developers call it an “everyday model” and compare it to GPT-5 in terms of response speed and response quality. The neural network, nicknamed “Speciale,” boasts “cutting-edge” reasoning capabilities.
The AIME and HMMT tests test neural networks’ ability to solve mathematical problems, IMO-AnswerBench contains 400 Olympiad problems with verifiable answers, GPQA Diamond tests knowledge in physics, biology, and chemistry, and HLE (Humanity’s Last Exam) includes academic questions in mathematics, humanities, and natural sciences.
DeepSeek emphasizes that V3.2-Speciale’s remarkable capabilities require a significant investment of tokens (and therefore computing power), so the team is not yet ready to make it publicly available.
DeepSeek-V3.2 is already available in the app, on the website, and through the company’s API services. The advanced V3.2-Speciale, currently only accessible via the API, offers maximum reasoning power for complex tasks.
