On August 21, DeepSeek unveiled its latest model, DeepSeek-V3.1, as open-source software for the first time on ModelScope. This latest iteration introduces three pivotal enhancements: Firstly, the model innovatively integrates both thinking and non-thinking modes, allowing users to seamlessly toggle between them via a simple click on the "Deep Thinking" button within the web interface or mobile application. Secondly, DeepSeek-V3.1-Think boasts a substantial improvement in processing efficiency, delivering answers more swiftly compared to its predecessor, DeepSeek-R1-0528, while also reducing the output token count by 20% to 50%. Lastly, leveraging Post-Training optimization technology, DeepSeek-V3.1 has achieved remarkable performance gains in tool utilization and agent-based tasks.
