Chinese AI lab DeepSeek has unveiled two advanced AI models, DeepSeek-V3.2 and DeepSeek-V3.2-Speciale, which aim to enhance reasoning capabilities in tool-use and complex inference tasks. This launch, announced on X, showcases DeepSeek’s ongoing commitment to developing cutting-edge AI technologies.
Both models, along with a detailed technical report, have been released as open source on Hugging Face, facilitating broader accessibility and collaboration in the AI research community. The V3.2 model is positioned as the official successor to the previous V3.2-Exp, with availability across DeepSeek’s app, web interface, and API. In contrast, the Speciale variant is accessible only through a temporary API endpoint until December 15, 2025.
DeepSeek describes V3.2 as a “daily driver” capable of delivering performance on par with GPT-5, emphasizing a careful balance between inference efficiency and long-context performance. The company claims that the Speciale model, designed for high-end reasoning tasks, “rivals Gemini-3.0-Pro,” aiming for gold-level proficiency in various competitive benchmarks, including the IMO, CMO, and ICPC World Finals.
The introduction of these models marks a significant expansion of DeepSeek’s agent-training approach, supported by a newly developed synthetic dataset comprising over 1,800 environments and 85,000 complex instructions. V3.2 is noted as the first model to integrate reasoning capabilities directly with tool usage, allowing for structured reasoning to function both internally and in collaboration with external tools.
In tandem with these model releases, DeepSeek has also updated its API. The usage pattern for V3.2 remains consistent with its predecessor, while the Speciale model is priced equally but does not support tool calls. A standout feature of V3.2, termed “Thinking in Tool-Use,” aims to enhance the user experience by providing deeper insights into the reasoning process. Further details on this capability are available in the developer documentation provided by DeepSeek.
In addition to these launches, DeepSeek has introduced a new open-weight model, DeepSeekMath-V2, which demonstrates impressive theorem-proving abilities in mathematics. This model recently achieved gold-level scores at the International Mathematics Olympiad (IMO) 2025, underscoring the lab’s commitment to advancing AI applications in diverse fields.
As AI continues to evolve, the introduction of DeepSeek’s latest models not only highlights the competitive landscape of AI technologies but also sets a new benchmark for reasoning capabilities in the industry. With the capabilities of V3.2 and Speciale now available for public use, DeepSeek is positioning itself at the forefront of AI innovation, potentially influencing future advancements and applications in the field.
See also
OpenAI, Google Impose Daily Limits on AI Tools: Sora at 6 Videos, Nano Banana Pro at 2 Images
Kling AI Launches Kling O1, First Unified Multimodal Video Model for Seamless Content Creation
Penn State Students Combine Art, Writing, and AI for Creative Self-Discovery Project
Pollo AI Launches Video Generator, Transforming Images and Text into Dynamic Clips
Pixazo Launches Kling O1 API, a Unified Multi-Modal Engine for Advanced Visual Creation
