DeepSeek Starts Testing V4.1 Flash, New Model with Multimodal Support

DeepSeek announced today that it's testing an intermediate version of V4.1 Flash. The new model has a new structure and natively supports multimodal input, offering better performance, faster speed, and lower cost. Testers can keep their base_url and just set the model name to deepseek-v4.1-flash-expires-on-0910. Pricing stays the same as the current V4 Flash, with a limit of 20 concurrent requests per account.

In a feedback form, DeepSeek asks users if this version could fully replace the online V4 Pro. Earlier, the V4 Flash API went public in late July with improved agent abilities. On August 21, the Vision-Exp model was added, supporting image input for tasks like describing pictures, reading text in screenshots, and analyzing charts.