Two months ago, DeepSeek released its experimental model V3.2-Exp and confirmed that it was not inferior to V3.1-Terminus in various scenarios through user return visits, verifying the effectiveness of DSA's sparse attention mechanism. The latest official versions of DeepSeek-V3.2 and DeepSeek-V3.2-Speciale are now available, in which V3.2 has completely replaced V3.2-Exp on the official website, app, and API as the default service model, and the Speciale version is open as a temporary API service for community evaluation and research.
According to the official, the goal of V3.2 is to strike a balance between reasoning ability and output length, suitable for daily scenarios such as Q&A and general agents, and reach the level of GPT-5 on the public reasoning benchmark, which is only slightly lower than Gemini-3.0-Pro, and is more restrained in output length and computational overhead than Kimi-K2-Thinking. V3.2-Speciale greatly enhances its long-term thinking ability and integrates DeepSeek-Math-V2's theorem proof ability, and won gold medals in the IMO, CMO, ICPC World Finals and IOI 2025, with ICPC and IOI performances close to the second and tenth places among human players respectively. This model specializes in complex reasoning tasks, with significantly higher token consumption, and currently only supports thinking mode dialogue, not open to tool calls, and has a maximum output length of 128K.
In terms of agent applications, DeepSeek-V3.2 has become the first model to directly integrate "thinking" into tool calls, which can not only support multiple rounds of "thinking + calling tools" in thinking mode, but also supporting non-thinking mode calls. The official constructed large-scale agent training data by synthesizing more than 1,800 environments and 85,000 complex instructions, so that the model reached the leading level of the current open source model in the evaluation of various agents. V3.2-Speciale is accessed through a temporary endpoint base_url configuration, and the service is expected to end at 23:59 Beijing time on December 15, 2025.
FAQ
Q: What is the difference between the positioning of DeepSeek-V3.2 and V3.2-Speciale?
A: V3.2 is aimed at daily Q&A and general agents, emphasizing reasoning and cost balance. V3.2-Speciale pursues extreme reasoning ability, focusing on difficult math and programming tasks.
Q: Which model is currently being called on the web and app?
A: DeepSeek has upgraded the default model of the web, app, and standard APIs to the official version of DeepSeek-V3.2, and no longer uses V3.2-Exp.
Q: How can I access V3.2-Speciale and what are the limitations?
A: You need to set a specific base_url in the API to call, this version only supports thinking mode conversations, does not support tool calls, the maximum output length is 128K, and the service is open until December 15, 2025.
Q: What exactly does "tool call in thinking mode" mean?
A: V3.2 can insert tool calls multiple times during the inference process, and then continue to think about it in combination with the return results of the tool, which is suitable for complex and multi-step agent task scenarios.
Q: Are both models already open source?
A: The official V3.2 and V3.2-Speciale have been released on HuggingFace and ModelScope for community download, along with technical reports and thinking mode API documentation.
