PANews reported on August 21st that DeepSeek announced the release of the V3.1 model, which for the first time implements a hybrid "Think & Non-Think" reasoning mode, which users can switch between using the "DeepThink" button. V3.1-Think mode boasts faster reasoning than DeepSeek-R1-0528 and enhances tool usage and multi-step proxy task capabilities. The API adds support for 128KB context, compatible with the Anthropic API format and strict function calls. The V3.1 base model extends long context based on the 840B token. Open source weights are now available, and pricing will be adjusted on September 5th.