DeepSeek has launched its V3.1 model, introducing a novel hybrid "Think & Non-Think" reasoning mode accessible via the "DeepThink" button. This update offers faster reasoning capabilities compared to the previous DeepSeek-R1-0528 model, along with improved tool usage and multi-step proxy task handling. The V3.1 model supports a 128KB context, aligning with the Anthropic API format and strict function calls. Additionally, the model's base extends long context based on the 840B token, with open source weights now available. Pricing adjustments are set to take effect on September 5th.