DeepSeek, a Chinese AI startup, has launched two new reasoning models, DeepSeek-V3.2 and DeepSeek-V3.2-Speciale, that the company says rival the performance of OpenAI’s GPT-5 and Google’s Gemini 3 Pro on complex reasoning and coding tests.

DeepSeek-V3.2, now powering the company’s app, website, and API, is designed as a “daily driver” AI model combining high reasoning capability with computational efficiency. In contrast, DeepSeek-V3.2-Speciale targets maximum reasoning power, delivering gold-level results in major competitions like the International Mathematical Olympiad and ICPC World Finals in 2025. According to DeepSeek, the Speciale model outperformed GPT-5 and Gemini on mathematical benchmarks like the American Invitational Mathematics Examination, scoring 96.0 per cent.

The models feature DeepSeek Sparse Attention, a novel mechanism that reduces computational costs by 50-70 per cent when handling long text sequences. The firm emphasises that these models integrate “thinking” directly into tool use, enabling better performance when combined with external tools such as code execution and search engines.

DeepSeek's release strengthens China’s position in AI innovation amid intense competition with U.S.-based companies. Chinese open-source AI models have gained global traction, now accounting for 17 per cent of worldwide downloads, surpassing American counterparts for the first time. This momentum reflects China’s strategy to develop accessible, efficient AI models despite restrictions on advanced chips.

DeepSeek, founded in 2023 with backing from the quantitative hedge fund High-Flyer, joins giants like Alibaba and Baidu in accelerating AI research and deployment in China. The V3.2-Speciale model is available via API until mid-December to support community evaluation and research.