DeepSeek-V4.1 Flash: Pushing the Limits of KV Cache Compression

DeepSeek’s new V4.1 Flash isn’t just another tune-up. The technical report frames it as a push to squeeze KV cache as far as it can go, and that matters because long-horizon agent workflows keep stretching context while tool use piles on prefill cost. Once the cache gets large, memory in HBM and on SSD starts to pinch, and serving gets slower and pricier.

The model is built around that problem. DeepSeek-V4.1-Flash is a multimodal mixture-of-experts system with 552B parameters and native support for up to 1 million tokens of context. It uses a causal encoder-decoder design so the prefill stage only activates 8B parameters per token, while decoding activates 16B. The company says that, at the same sequence length, its runtime KV storage is about one quarter of DeepSeek-V4-Flash and its persistent KV storage is about one eighth.

A lot of the gain comes from architecture, not just a trick at the end. The report says the model has 40 layers, but only the first 20 are needed during prefill. On the cache side, DeepSeek combines several forms of compression: fewer heads, block-style compression, cross-layer reuse, and FP4 KV cache. The result is a further 4x compression of KV cache while keeping task quality high.

The attention design is the other half of the story. DeepSeek-V4.1-Flash uses a sparse setup with a sliding window for local context and a global cache for long-range reuse. It also trims the work of sparse attention’s indexer. The paper describes a 512-dimensional latent for shared key-value representation, encoder-side merging of adjacent positions, and multiple layers sharing the same global KV.

There’s also multimodal plumbing here. The vision path uses a 32-layer ViT, then compresses image features before inserting them into the language stream. And the whole point of all this machinery is practical: keep long-context serving from turning into a memory and bandwidth tax. That, more than the flashy token speed, is the real headline.