[DeepSeek V3.2 Official Release: Enhanced Agent Capabilities, Integrated Thinking and Reasoning] Reported by Jinse Finance, today we are releasing two official model versions simultaneously: DeepSeek-V3.2 and DeepSeek-V3.2-Speciale. DeepSeek-V3.2 is our first model that integrates thinking into tool usage, supporting both thinking mode and non-thinking mode for tool invocation. We have proposed a large-scale Agent training data synthesis method, constructing numerous 'hard-to-answer, easy-to-verify' reinforcement learning tasks (1800+ environments, 85,000+ complex instructions), significantly improving the model's generalization capabilities. (DeepSeek)