DeepSeek Releases V4.1 Flash Model with Native Multimodal Support and Architecture Innovations
DeepSeek has officially released DeepSeek V4.1 Flash, a 552-billion parameter Mixture-of-Experts (MoE) model featuring native multimodal vision understanding. It outperforms the previous flagship V4 Pro across performance benchmarks, speed, and cost, prompting a lower API pricing structure with off-peak discounts. The release sets a new standard for serving large-scale multimodal models by drastically lowering memory requirements and operational costs. By significantly compressing Key-Value (KV) cache overhead, long-context processing and complex AI agent workflows become substantially more affordable for developers and enterprise applications. Built on a novel asymmetric Causal-Encoder-Decoder architecture, V4.1 Flash activates 8B parameters for input encoding and 16B for output generation out of its total 552B parameters. Compared to the previous generation, its KV cache reduces High Bandwidth Memory (HBM) usage by 75% and SSD storage by 87.5%, achieving a cumulative 437-fold reduction compared to DeepSeek's first-generation model.
## BACKGROUND
Large language models rely on Key-Value (KV) caching to store attention states for fast token generation, but this consumes enormous High Bandwidth Memory (HBM) during long-context inference. Additionally, Mixture-of-Experts (MoE) architectures allow models to scale up total parameters while only activating a subset of parameters per request to maintain inference speed.