New AI Model DeepSeek-V3.2 Released

DeepSeek has announced two new AI models — DeepSeek-V3.2 and DeepSeek-V3.2-Speciale. The primary focus of these models is on advanced reasoning capabilities, enabling them to tackle problems at the level of international student olympiads.

According to the developers, the special V3.2-Speciale version delivers performance comparable to the latest flagship models from OpenAI and Google. For instance, on the American Invitational Mathematics Examination (AIME), the model achieved a score of 96%, slightly ahead of GPT-5 High (94.6%) and Gemini 3 Pro (95%). In programming tests, the gap is minimal — all three models fall within the 73-76% range.

A key feature of the new models is their architecture, optimized to function as AI agents. DeepSeek has, for the first time, directly integrated "reasoning" into the tool-usage process, allowing the neural network to plan actions for solving multi-step problems. Training utilized synthetic data covering over 1,800 different environments.

Users can already test the base V3.2 version through the official app, web interface, or API. It is positioned as an everyday tool with performance close to GPT-5, while the more powerful V3.2-Speciale is currently available only via a special API endpoint until mid-December 2025, as it requires significant computational resources. Both models are distributed with open-source code on the Hugging Face platform.

The release of the new models comes just two months after the experimental V3.2-Exp version, demonstrating the rapid development pace in this field. Experts note that such achievements strengthen China's position as a leader in creating sophisticated artificial intelligence systems.