DeepSeek Unveils New AI Models Aiming to Challenge Google and OpenAI

China’s DeepSeek has rolled out two upgraded versions of the experimental artificial-intelligence model it introduced only weeks ago, adding new capabilities the startup says will strengthen both reasoning and autonomous task execution.

‍

The company’s earlier release in September carried the name DeepSeek-V3.2-Exp to emphasize its experimental status. The newly announced update, simplified to DeepSeek-V3.2, is being positioned as a major step forward.

‍

According to the Hangzhou-based startup, the latest version delivers performance on par with OpenAI’s flagship GPT-5 across a wide range of reasoning benchmarks an indication that China’s open-source AI ecosystem remains competitive with Silicon Valley’s frontier, closed-source models in several key areas.

‍

DeepSeek also highlighted that V3.2 goes beyond pure reasoning by integrating tool-use capabilities. The model can operate search engines, calculators, and code execution tools, fusing analytical thinking with real-time actions. “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 company said in a post on X. The original experimental version was described as an early preview of DeepSeek’s ambitions for next-generation AI systems.

‍

Alongside the main release, the company introduced a second variant, DeepSeek-V3.2-Speciale, which is tailored for mathematical reasoning and other forms of extended, complex problem solving. DeepSeek said the goal of Speciale is to “push the inference capabilities of open-source models to their limits and explore the boundaries of model capabilities.”

‍

According to the startup, the Speciale version matches the performance of Google’s latest Gemini-3 Pro and scored at gold-medal levels on elite global competitions such as the International Math Olympiad and the International Olympiad in Informatics.

‍

DeepSeek also revealed that it has developed a new methodology for training AI agents systems designed to operate autonomously, adapt to their surroundings, analyze information, and make decisions with minimal human intervention. The approach is intended to improve how these agents break down tasks, manage long-horizon reasoning, and execute actions efficiently in real-world environments.

‍

Taken together, the releases signal that the influential Chinese AI lab is accelerating its research momentum and strengthening its position in the country’s rapidly evolving AI race. After capturing the global tech community’s attention in January with a model that reshaped expectations for open-source AI performance, DeepSeek continues to push ahead with innovations designed to boost efficiency, expand model capabilities, and elevate the overall speed of AI processing.

‍

The latest models also arrive shortly after DeepSeek introduced DeepSeekMath-V2, an open AI model with advanced theorem-proving abilities. The company’s recent cadence of releases underscores a broader strategy: to build a portfolio of highly specialized and high-performing models that can challenge or even surpass the capabilities of proprietary systems developed by US-based tech giants.

‍

In a detailed technical report titled “DeepSeek-V3.2: Pushing the Frontier of Open Large Language Models,” the startup reiterated that DeepSeek-V3.2 reaches performance levels comparable to Kimi-k2-thinking and GPT-5 on multiple reasoning tests. This claim, if validated, would represent a significant milestone for China’s AI ambitions, demonstrating that open-source models can hold their own against the most advanced commercial offerings.

‍

As competition between global AI labs intensifies, DeepSeek’s latest releases reflect a broader trend toward models that combine deep reasoning skills with real-world tool interoperability an area many researchers believe will define the next decade of AI development. With V3.2 and V3.2-Speciale, DeepSeek is signaling that it intends not only to keep pace with the world’s leading labs, but to shape the direction of the next wave of open-source AI innovation.

‍