DeepSeek releases two new models

On December 1st, DeepSeek announced the release of two official models: DeepSeek-V3.2 and DeepSeek-V3.2-Speciale. The official website, app, and API have all been updated to the official version DeepSeek-V3.2, so feel free to use it. The Speciale version is currently only available in the form of temporary API services for community evaluation and research purposes. World-leading reasoning capabilities The goal of DeepSeek-V3.2 is to balance reasoning capabilities with output length, making it suitable for daily use in scenarios such as question-answering and general agent tasks. In public reasoning benchmarks, DeepSeek-V3.2 has reached the level of GPT-5, slightly below Gemini-3.0-Pro; compared to Kimi-K2-Thinking, V3.2 has significantly reduced output length, thus reducing computation costs and user wait times. The goal of DeepSeek-V3.2-Speciale is to push the reasoning capabilities of open-source models to the extreme and explore the boundaries of model capabilities. V3.2-Speciale is an enhanced version of DeepSeek-V3.2 for deep thinking, combining the theorem proving capabilities of DeepSeek-Math-V2. This model excels in instruction following, rigorous mathematical proof, and logical verification capabilities, performing on par with Gemini-3.0-Pro in mainstream reasoning benchmarks. More notably, the V3.2-Speciale model has successfully won gold medals in IMO 2025 (International Mathematical Olympiad), CMO 2025 (Chinese Mathematical Olympiad), ICPC World Finals 2025 (International Collegiate Programming Contest World Finals), and IOI 2025 (International Olympiad in Informatics). The ICPC and IOI scores reached the levels of second and tenth place among human participants, respectively. Tips: In highly complex tasks, the Speciale model significantly outperforms the standard version, but also consumes significantly more tokens and has higher costs. Currently, DeepSeek-V3.2-Speciale is intended for research use only and does not support tool calls. It has not yet been specifically optimized for daily conversations and writing tasks. Incorporating thinking into tool calls Unlike previous versions that were limited in the ability to call tools while in thinking mode, DeepSeek-V3.2 is the company's first model to incorporate thinking into tool usage and simultaneously support tool calls in both thinking and non-thinking modes. The company has proposed a large-scale Agent training data synthesis method, creating a large number of "difficult to answer, easy to verify" reinforcement learning tasks (1800+ environments, 85,000+ complex instructions), significantly improving the model's generalization ability. As shown in Table 2, the DeepSeek-V3.2 model achieves the highest level among current open-source models in intelligent agent evaluations, greatly narrowing the gap between open-source and closed-source models. It is worth noting that V3.2 has not undergone special training for these test sets, so we believe that V3.2 will demonstrate strong generalization capabilities in real-world application scenarios.