China’s DeepSeek has unveiled two new versions of its fast-evolving artificial-intelligence model, escalating the global contest over advanced reasoning systems. Just weeks after releasing an experimental build dubbed DeepSeek-V3.2-Exp, the Hangzhou-based company has launched its official successor, DeepSeek-V3.2, alongside a higher-powered edition called V3.2-Speciale.
DeepSeek-V3.2 is positioned as a full-strength upgrade, with the company asserting that it matches OpenAI’s flagship GPT-5 across a range of reasoning benchmarks. The startup says the model blends human-style logical processing with the capacity to autonomously use tools such as search engines, calculators, and code executors. In a post on X, the company described it as the first DeepSeek system that “integrates thinking directly into tool-use,” enabling both reflective and fast-execution modes in agent-style workflows.
The release underscores DeepSeek’s rapid rise after shocking the industry in January with a breakthrough model that briefly rattled global AI valuations. The new version aims to sharpen the competitive edge of China’s open-source ecosystem, signaling that local innovators remain capable of challenging Silicon Valley’s proprietary giants on key technical fronts.
Alongside the main release, DeepSeek rolled out V3.2-Speciale, a variant built specifically for intense mathematical and long-form reasoning tasks. The company says Speciale reaches parity with Google’s latest Gemini-3 Pro and performs at gold-medal levels on elite competitions such as the International Math Olympiad and the International Olympiad in Informatics. According to DeepSeek, the version is designed to push the upper limits of inference efficiency while testing how far open-source models can be stretched.
The startup also claims to have developed a new training method for autonomous AI agents—software systems capable of interacting with their environment, analyzing information, and making decisions independently of human operators. The approach is intended to speed up model reasoning while reducing computational overhead, an area where the company has been racing ahead following its disruptive January breakthrough.
The announcement follows a series of rapid-fire releases, including last week’s DeepSeekMath-V2, an open model praised for advanced theorem-proving skills. In an accompanying technical report titled DeepSeek-V3.2: Pushing the Frontier of Open Large Language Models, the company argues that V3.2 delivers performance on par with Kimi-k2-thinking and GPT-5 across multiple reasoning evaluations—further evidence, it says, that open systems can hold their ground against the world’s most advanced proprietary AI.
