On the evening of December 1, DeepSeek unveiled its latest-generation open-source large-scale model, DeepSeek-V3.2, along with an enhanced long-thinking iteration, DeepSeek-V3.2-Speciale. Notably, DeepSeek-V3.2 retains exceptional reasoning prowess while striking an optimal balance in computational efficiency, rendering it ideal for everyday Q&A scenarios. In public reasoning evaluations, it has demonstrated performance on par with GPT-5, albeit slightly trailing behind Gemini-3.0-Pro. Its shorter output lengths effectively minimize computational demands and reduce user waiting times. Conversely, DeepSeek-V3.2-Speciale integrates the theorem-proving capabilities of DeepSeek-Math-V2, boasting remarkable instruction adherence, rigorous mathematical proofing, and logical verification skills. This variant has clinched gold medals in numerous international contests and is presently accessible solely for research endeavors. The official web portal, mobile application, and API have all been upgraded to the V3.2 version.