DeepSeek V4.1 Flash Officially Released, Reducing Demand for HBM and SSDs

DeepSeek V4.1 Flash model was recently officially released. It is the smallest model in DeepSeek's new model architecture series and features native multimodal visual understanding capability.

DeepSeek V4.1 Flash is a 552B-parameter MoE model that adopts new Causal-Encoder-Decoder architecture, with asymmetric input and output that only 8B activated on the input side and 16B activated on the output side, resulting in significantly lower cost than known models of the same size. At the same time, V4.1 Flash also adopts a new pre-training approach and has undergone larger-scale reinforcement learning post-training. In benchmark tests, it successfully surpassed the intelligence level of a range of flagship models, including DeepSeek V4 Pro.

Notably, DeepSeek V4.1 Flash significantly reduces the size of the KV Cache. Compared with the previous generation model, its demand for HBM has been reduced to 1/4, and its demand for SSDs has been reduced to 1/8. In Agent usage scenarios, cache hit costs often account for a relatively high proportion. The change in DeepSeek V4.1 Flash's KV Cache requirement means that the same task requires less high-speed memory and storage resources, substantially lowering the usage cost of Agent task.