Chinese company DeepSeek has introduced two new versions of its experimental AI model that was launched a few weeks ago. According to the startup, new capabilities have been added that help combine thinking and independent execution of certain actions.
While the model released in September was called DeepSeek-V3.2-Exp — a nod to the experimental nature — the new version is simply called DeepSeek-V3.2. The startup claims that the new service matches the performance of GPT-5, OpenAI’s flagship model, in several inference tests, suggesting that open source systems developed in China remain competitive with cutting-edge Silicon Valley models on at least some indicators.
DeepSeek also stated that V3.2 combines the ability to simulate something close to human thinking with tools such as search engines, calculators, and code executors.
“DeepSeek-V3.2 is our first model that integrates the thinking process directly into tool use, and also supports tool use in both thinking and non-thinking modes,” the company said in an X post.
DeepSeek, which took the tech sector by storm in January with a groundbreaking AI model, had previously described the beta as a step toward the next generation of technology.
The second version released, V3.2-Speciale, focuses on mathematical calculations and other abilities that require long thinking. According to the startup, the goal is to “push the inference capabilities of open source models to the limit and explore the limits of what these models can do.”
DeepSeek-V3.2-Speciale matches the performance of Gemini-3 Pro, Google’s latest model, and has achieved gold medal levels in benchmark tests such as the International Mathematical Olympiad and the International Computer Olympiad, according to the company.
DeepSeek said it has created a new way to train artificial intelligence agents, which are programs designed to work autonomously, interact with the environment, analyze data, and make decisions without constant human supervision.
It’s another sign that the influential lab is ramping up its research momentum to make AI processing faster and more efficient. After its launch at the beginning of the year, the startup is now moving forward with innovations to strengthen its advantage in the Chinese AI race.
Last week, the company launched a new open model, DeepSeekMath-V2, which has demonstrated strong theorem proof capabilities.
“DeepSeek-V3.2 achieves similar performance to Kimi-k2-thinking and GPT-5 on several inference benchmarks,” the startup said in a technical report accompanying the release titled. “DeepSeek-V3.2: Pushing the boundaries of large open language models“.