DeepSeek Begins Closed Beta Testing for DeepSeek V4.1 Flash Model
DeepSeek has launched a closed beta for DeepSeek V4.1 Flash, an intermediate release built on a redesigned model architecture with native multimodal support. The updated model offers enhanced performance, faster inference speeds, and lower operational costs. This release demonstrates DeepSeek's rapid iteration cycle in delivering native multimodal reasoning within a lightweight, cost-efficient Flash architecture. DeepSeek is actively gathering feedback to evaluate whether this optimized Flash variant can fully replace its flagship online V4 Pro model. Beta testers can invoke the model using the identifier `deepseek-v4.1-flash-expires-on-0910` while retaining their existing base API endpoint. The model shares the same billing rates as DeepSeek V4 Flash and enforces a rate limit of 20 concurrent connections per account.
## BACKGROUND
Native multimodal architectures are built from scratch to process and combine disparate data types—such as text, vision, and audio—within a single neural network rather than linking separate vision encoders and text models together. DeepSeek previously launched its V4 Flash API in late July with enhanced AI agent functionality, followed by an experimental multimodal vision model in August.