China’s DeepSeek has introduced two upgraded versions of an experimental artificial intelligence model it released only weeks earlier, expanding its capabilities in reasoning and autonomous task execution.
The startup said the additions mark a further step in enhancing how AI handles complex thinking and practical tool use.
The model released in September, known as DeepSeek-V3.2-Exp, has now been refined into a new version called DeepSeek-V3.2.
The company said this updated model matches OpenAI’s GPT-5 on several reasoning benchmarks, highlighting that China’s open-source systems can still compete with Silicon Valley’s top proprietary models on specific metrics.
DeepSeek said the V3.2 model blends reasoning abilities with operational tools, including search engines, calculators, and code executors.
In a post on X, the company stated, “DeepSeek-V3.2 is our first model to integrate thinking directly into tool-use, and also supports tool-use in both thinking and non-thinking modes.”
The startup had earlier described its experimental version as part of its path toward next-generation AI. DeepSeek gained global attention in January after releasing a breakthrough model that surprised the tech industry.
Alongside V3.2, the company unveiled a second model called DeepSeek-V3.2-Speciale. This version is designed for advanced mathematical reasoning and sustained long-form thinking.
DeepSeek said the aim is to “push the inference capabilities of open-source models to their limits and explore the boundaries of model capabilities.” According to the company, V3.2-Speciale performs on par with Google’s Gemini-3 Pro and reaches gold-medal-level scores in competitions such as the International Math Olympiad and the International Olympiad on Informatics.
DeepSeek also announced that it has developed a new method for training AI agents — systems that can act independently, assess data, interact with their surroundings, and make decisions without constant human involvement. The company said these innovations indicate its commitment to improving AI efficiency and speed.
Following a series of major advancements in January, DeepSeek has continued intensifying its research efforts. Just last week, it released DeepSeekMath-V2, an open model with strong mathematical theorem-proving abilities.
In a technical report titled “DeepSeek-V3.2: Pushing the Frontier of Open Large Language Models,” the company said, “DeepSeek-V3.2 achieves similar performance with Kimi-k2-thinking and GPT-5 across multiple reasoning benchmarks.”