DeepSeek has introduced two new open-source artificial intelligence models, V3.2 and V3.2-Speciale, signaling a renewed effort to compete with major players in the global AI race. The announcement follows the company’s sharp rise in early 2025, when its earlier breakthrough models rattled U.S. stock markets and drew industry-wide attention.
The company says the standard V3.2 model is designed as a balanced, general-purpose system and is positioned as a “daily driver” for broad use. In its announcement, DeepSeek stated, “The standard DeepSeek-V3.2 serves as a daily driver that harmonizes efficiency with agentic performance on par with GPT-5.” The model is also the first in DeepSeek’s lineup to integrate built-in “thinking” directly into tool-use operations.
See also: Survey Shows Majority of Germans Now Use Generative AI in Daily Life
More emphasis, however, has been placed on the high-compute V3.2-Speciale variant. DeepSeek claims the model surpasses GPT-5 and rivals Google’s Gemini 3 Pro in pure reasoning capability, citing gold-medal-level results at the 2025 International Mathematical Olympiad and the International Olympiad in Informatics. To reinforce transparency, the company said, “We have published all final submissions for public verification.”
According to DeepSeek, the performance improvements are driven by its internal DeepSeek Sparse Attention (DSA) mechanism, which reduces computational overhead in long-context tasks, as well as a large-scale reinforcement learning framework. The company also highlighted a “Large-Scale Agentic Task Synthesis Pipeline” used to train the models on more than 85,000 complex instructions.
See also: AI Notification Tools Lead Android 16 Upgrade Push
The standard V3.2 model is currently available via web, mobile and API platforms, while V3.2-Speciale remains restricted to API access and operates through a temporary endpoint expiring on December 15, 2025. Speciale functions solely as a reasoning engine and does not support tool calling. DeepSeek has also released instructions for developers wishing to run the models locally, reinforcing its open-source positioning in the competitive AI ecosystem.