DeepSeek Launches V3.1 Model with Hybrid Think and Non-Think Modes
DeepSeek releases V3.1, introducing a hybrid inference system that lets users toggle between thinking and non-thinking modes. The update brings stronger agent capabilities, Anthropic API support, and new open-source weights.
DeepSeek unveils DeepSeek-V3.1 as the company's first major step into the agent era, headlined by a new hybrid inference system. This single model operates in two distinct modes—Think and Non-Think—allowing users to easily toggle between standard chat responses and deeper reasoning via the "DeepThink" button. The thinking mode reaches conclusions significantly faster than the previous DeepSeek-R1-0528 model.
The update brings substantial upgrades for developers using the DeepSeek API, mapping the deepseek-chat endpoint to non-thinking mode and the deepseek-reasoner endpoint to thinking mode, both supporting a 128K context window. Developers now benefit from Anthropic API format compatibility and a beta release of Strict Function Calling, which dramatically improves the model's performance on complex tool-use and multi-step agent tasks like SWE and Terminal-Bench.
Under the hood, DeepSeek builds the V3.1 Base model on top of V3 using 840 billion tokens of continued pre-training specifically for long context extension, while also updating the tokenizer and chat template. The base and instruct model weights are available as open-source on Hugging Face, and users have until September 5, 2025, to utilize current API pricing before the newly announced rates take effect.