DeepSeek releases V3.2 models with GPT-5-level reasoning and gold medals in global competitions
On December 1, 2025, Chinese AI research group DeepSeek officially released two new open-source models: DeepSeek-V3.2 and DeepSeek-V3.2-Speciale. The standard V3.2 model is designed for everyday use, while the Speciale version targets advanced reasoning tasks and has already earned top honors in international competitions.
According to DeepSeek, the V3.2 model matches GPT-5 in public reasoning benchmarks and delivers faster output with shorter responses than Kimi-K2-Thinking, making it more efficient for practical applications.
The Speciale variant integrates theorem-proving capabilities from DeepSeek-Math-V2 and is optimized for long-form reasoning. It has won gold medals in four major events: the International Mathematical Olympiad (IMO 2025), Chinese Mathematical Olympiad (CMO 2025), International Olympiad in Informatics (IOI 2025), and the International Collegiate Programming Contest (ICPC). In the ICPC and IOI, its scores ranked second and tenth among human participants.
Both models are now open-sourced on HuggingFace and ModelScope. The Speciale version is temporarily available via API until December 15 for community testing and research.
DeepSeek is a Chinese AI research organization focused on developing open-source large language models and advanced reasoning systems. The team specializes in high-performance model architecture and mathematical logic integration. DeepSeek has gained recognition for its competitive performance in global benchmarks and academic contests. Its models are designed for both public use and enterprise-level deployment.
