On the night of December 1, DeepSeek unveiled two fresh models: DeepSeek - V3.2 and DeepSeek - V3.2 - Speciale. Among these, V3.2 is crafted to strike a balance between reasoning abilities and output length, rendering it well - suited for everyday Q&A sessions and common Agent task scenarios. In public reasoning tests, it has attained performance on par with GPT - 5. Although its performance is slightly inferior to that of Gemini - 3.0 - Pro, it significantly cuts down on output length. This reduction not only lowers computational expenses but also shortens the waiting time for users.

V3.2 - Speciale serves as an upgraded version of V3.2, boasting extended thinking capabilities. It integrates the theorem - proving prowess of DeepSeek - Math - V2. Its primary objective is to push the reasoning capabilities of open - source models to their absolute limits and delve into the boundaries of model capabilities. This model has clinched gold medals in numerous international competitions. At present, the official web interface, App, and API have all been upgraded to the official V3.2 version. Meanwhile, the Speciale version is currently accessible only as a temporary API service, intended for community evaluation and research purposes.