DeepSeek Begins API Beta Testing for DeepSeek V4.1 Flash
DeepSeek has launched internal API beta testing for DeepSeek V4.1 Flash, accessible using the model identifier `deepseek-v4.1-flash-expires-on-0910`. The update introduces a new model architecture with native multimodal capabilities designed for enhanced performance, higher inference speeds, and lower operational costs. As DeepSeek continues to compete aggressively on AI cost-efficiency, transitioning to a native multimodal architecture could significantly reduce latency and operational expenses for complex tasks. This release indicates rapid iteration toward DeepSeek's upcoming V4 model lineup. During the beta testing period, pricing remains identical to `deepseek-v4-flash`, with a concurrency rate limit of 20 requests per account. Developers can access the preview model without modifying their base API URL.
## BACKGROUND
Traditional multimodal models typically connect separate vision encoders to a language model backbone using modular adapter layers. In contrast, native multimodal architectures process multiple input modalities directly within a single unified pipeline, improving model efficiency and integration across tasks.