DeepSeek has launched DeepSeek-V3.1, marking a significant step forward toward what it calls the “agent era” in artificial intelligence. The new model features a hybrid inference system called “Think & Non-Think,” allowing one model to operate in two modes depending on the task.
DeepSeek-V3.1-Think, the thinking mode, delivers answers faster than the previous version DeepSeek-R1-0528. This update also includes post-training enhancements that improve the model’s ability to use tools and complete multi-step agent tasks more effectively.
Users can now toggle between the Think and Non-Think modes through the new “DeepThink” button available at DeepSeek’s chat platform.
Additional updates include API improvements, support for Anthropic API format, and Beta API with strict function calling. The model was trained on a massive 840 billion tokens to extend its long-context capabilities. Open-source weights and new tokenizer configurations are also available.
DeepSeek has announced pricing changes will take effect on September 5, 2025, ending the current off-peak discounts. Existing API pricing remains in place until that date.
