On December 2nd, according to the official DeepSeek WeChat account, versions V3.2 and V3.2-Speciale have been officially released. V3.2 achieves GPT-5 level performance on inference benchmarks, supports tool calls in thinking mode, and significantly improves agent generalization ability. V3.2-Speciale focuses more on mathematical and programming scenarios, achieving top-ten human performance in competitions such as IMO and IOI. The web, app, and API have been upgraded; the Speciale model API is temporarily open until December 15th.