PANews, August 21 news, DeepSeek announced the launch of the V3.1 model, which for the first time realizes the "Think & Non-Think" hybrid reasoning mode, allowing users to switch via the "DeepThink" button. The reasoning speed of the V3.1-Think mode has improved compared to DeepSeek-R1-0528, and it enhances tool usage and multi-step agent task capabilities. The API has added 128K context support, compatible with Anthropic API format and strict function calls. The V3.1 base model is based on 840B tokens to extend long contexts, and open-source weights have been released, with pricing to be adjusted on September 5.