DeepSeek releases two new AI reasoning models to compete with OpenAI and Google

Chinese AI startup DeepSeek has released two new AI models called DeepSeek-V3.2 and DeepSeek-V3.2-Speciale, on December 1. The startup has claimed that DeepSeek-V3.2 is designed as a daily AI assistant for reasoning on par with GPT-5, while DeepSeek-V3.2-Speciale is for more complex tasks and is comparable with the Gemini-3.0-Pro. 

The latter has achieved gold-medal performance in four elite international competitions – the 2025 International Mathematical Olympiad, the International Olympiad in Informatics, the ICPC World Finals and the China Mathematical Olympiad.

DeepSeek-V3.2 has also introduced an ability called “thinking in tool-use,” that enables them to reason through a problem while performing tasks like coding, web searching and file editing. The AI model also supports tool-use in both thinking and non-thinking modes.

The researchers trained the AI models using a new method called DeepSeek Sparse Attention or DSA that lessens high compute costs while maintaining performance.

The firm has also released the full technical report for the training process for both models. DeepSeek-V3.2 is now live on the app, web and via API with DeepSeek-V.3.2-Speciale just available via API.