DeepSeek has released V4.1 Flash, a 552-billion-parameter model with native image understanding that the company says exceeds V4 Pro in its official benchmarks. The model uses a Causal-Encoder-Decoder architecture that separates input processing from response generation, activating 8 billion parameters when reading inputs and 16 billion when producing outputs. DeepSeek said the design reduces inference costs despite the model's total size. V4.1 Flash also cuts high-bandwidth memory, or HBM, requirements to one-quarter of the prior generation's level and SSD storage needs to one-eighth, targeting lower costs for long-context workloads and repeated agent calls. Developers can access it through DeepSeek's API using the model name deepseek-flash. DeepSeek is still scheduled to route V4 Pro requests to V4.1 Flash at Flash pricing at 12:00 on Sept. 14, after delaying an initially planned immediate migration following developer concerns over production compatibility.