DeepSeek-v3.2 Release (api-docs.deepseek.com)

DeepSeek has announced the release of DeepSeek-V3.2 and DeepSeek-V3.2-Speciale, the latest iterations of its AI reasoning models designed for agents. The V3.2 version, now accessible on App, Web, and API, offers balanced inference and performance comparable to GPT-5, making it a versatile tool for daily AI tasks. Meanwhile, the V3.2-Speciale is touted for its enhanced reasoning capabilities that rival Gemini-3.0-Pro, achieving gold medal performance in prestigious competitions such as the IMO and IOI 2025. Though currently API-only and requiring more tokens, this model is tailored for complex tasks and supports a new approach to agent training. Significantly, DeepSeek-V3.2 is the first to integrate thinking directly into tool-use, enhancing its functionality across diverse environments with over 85,000 complex directives. This development introduces a synthesis method for massive agent training data covering more than 1,800 scenarios. Both models will also maintain similar API usage patterns, and they are open-sourced on Hugging Face, promoting community evaluation and further research. As it stands, DeepSeek’s innovations may reshape the landscape of AI/ML applications, offering sophisticated reasoning mechanisms that challenge existing paradigms.