DeepSeek has formally released DeepSeek V4.1 Flash, a new artificial-intelligence model that the company says is built for faster inference, higher throughput and native visual understanding. The announcement was dated September 10, 2026, and puts the model in the company’s API immediately.
DeepSeek describes V4.1 Flash as a 552-billion-parameter mixture-of-experts model. Its design uses an asymmetric causal-encoder-decoder structure, with eight billion parameters active for input processing and 16 billion active during output generation. Those figures describe the architecture, not a guarantee of response quality for every task.
The company said the model was trained with a larger reinforcement-learning phase and performed above DeepSeek V4 Pro in its own benchmark comparisons. That claim is DeepSeek’s, so the result should be read as a company-reported measurement rather than an independent ranking of all leading models.
V4.1 Flash is available through the DeepSeek API with native multimodal support. Developers can call the current model by using the deepseek-flash name. DeepSeek also said the older V4 Flash and V4 Flash Vision experimental names will be routed to the new model for compatibility.
The release changes how DeepSeek plans to handle its Pro service. From noon Beijing time on September 14, requests sent to deepseek-v4-pro will be routed to V4.1 Flash until a future V4.1 Pro release. Those requests will be billed at the Flash-series price, according to the company.
DeepSeek also lowered V4.1 Flash API prices and kept a peak and off-peak structure. The company said the new rates took effect at noon Beijing time on September 10, with off-peak prices set at half the peak rates to encourage users to move workloads when demand is lower.
The company named WorkBuddy, CodeBuddy and OpenCode as partners that have integrated V4.1 Flash. It also said it plans to support open-source inference work and pointed larger deployment teams toward the model release and technical report. The announcement gives developers a new option, but independent testing will still be needed before broad performance comparisons are made.
