DeepSeek released V4.1 Flash, the first and smallest model in a new architecture family, with native visual understanding and asymmetric activation for input and output. The company will route V4 Pro requests to Flash from September 14 until a successor arrives, making the faster model the temporary default rather than merely adding another option.
DeepSeek released V4.1 Flash as the first and smallest member of a new model architecture family. The company describes it as a mixture-of-experts model with 552 billion total parameters: 8 billion are active while processing input and 16 billion while producing output. Native visual understanding is part of the design. TechNode independently reported the formal September 10 launch and the model's place in the Flash line.[1], [2]
The release also changes what existing API customers receive. DeepSeek says requests addressed to V4 Pro will be routed to V4.1 Flash after noon Beijing time on September 14 and will remain there until the future V4.1 Pro is released. TechNode reports the same transition. Requests using the older V4 Flash and V4 Flash Vision Experimental names are also being redirected temporarily, consolidating several access paths around V4.1 Flash.[1], [2]
DeepSeek says the asymmetric activation pattern is intended to increase serving speed and throughput while lowering the compute used for each request. New API prices took effect early on September 10, with off-peak rates set below peak-hour rates. These efficiency and capability statements remain the company's own claims; the retained reports establish the release and routing change, not an independent benchmark result showing that Flash outperforms competing systems.[1], [2]
