DeepSeek-v4.1-Flash: Pushing the Limits of KV Cache Compression (arxiv.org)

DeepSeek-V4.1-Flash has been introduced as a groundbreaking multimodal Mixture-of-Experts (MoE) model designed to optimize KV cache compression, targeting the increasing computational demands of long-horizon agents. With a staggering 552 billion backbone parameters and the capability to handle contexts up to one million tokens, this model addresses the cost inefficiencies associated with long-context computations. The innovative Causal Encoder-Decoder architecture allows for significant parameter activation reduction—activating 16 billion during decoding and only 8 billion during prefill—thus enhancing cost efficiency for agentic workloads. Significantly, DeepSeek-V4.1-Flash advances KV cache compression through its implementation of compressed sparse attention and FP4 KV caching, leading to a global KV cache footprint of just 890 bytes per token, approximately one-fourth of its predecessor's size. This reduction is further enhanced by the SWA Bounded Replay optimization, cutting the persistent KV cache footprint to about one-eighth of that of the previous model. The extensive pretraining on a multimodal corpus of 45 trillion tokens, along with post-training refinements, enables DeepSeek-V4.1-Flash to outperform previous models across various text and multimodal tasks, marking a significant milestone in the evolution of large-scale AI computation.