Glonce

DeepSeek-V4.1-Flash compresses KV cache 4x with new CED architecture

modelsThe Glonce editorsHacker News🔥78Glonce rating78Impact65Novelty70Relevance100Surprise45Players85Credibility75

A blog post by an outside analyst works through DeepSeek's newly released technical report, "DeepSeek-V4.1-Flash: Pushing the Limits of KV Cache Compression," and argues that despite the "V4.1" name, the changes are big enough that the release deserves to be treated as a DeepSeek-V5. The author writes that after using the model for a while, they found it running at nearly 420 tokens per second, and that a person referred to as Cui said all DeepSeek-V4 Pro models would be taken offline, two signals that led the author to treat this as more than a routine post-training update. This piece is the first in a planned series covering the report and focuses on the model architecture chapter by chapter; the author notes that more critical material is still to come in Chapter 3.

DeepSeek-V4.1-Flash is a 552B-parameter multimodal mixture-of-experts model that natively handles multimodal input and supports contexts of up to 1 million tokens. Its central design is a Causal Encoder-Decoder (CED) architecture: the Decoder's global key-value cache is produced by projecting the Encoder's final hidden states, so most of a long prompt only has to pass through the first 20 of the model's 40 layers. As a result, the model activates just 8B parameters per token during the Prefill stage and 16B parameters per token during Decoding, a split the author says is especially cost-effective for input-heavy agent scenarios.

The report frames this work as a response to Long-horizon Agent Workflows, which push context windows longer and longer while tool calls add heavy Prefill computation; the resulting KV cache then has to be stored in HBM and on external SSDs and moved around, which the author says makes further scaling impractical without compression. DeepSeek's answer combines several techniques under a new attention variant called CSA2, compressing the KV cache along three dimensions: a 512-dimensional latent vector shared across the keys and values of each attention head (channel dimension), merging of 2 adjacent positions into 1 cache entry inside the Encoder (sequence dimension), and cross-layer sharing so the whole network keeps only 3 Encoder cache copies and 1 Decoder cache copy (layer dimension). Combined with FP4-precision KV cache quantization, storage growth for the global main KV plus its indexer comes to about 890 bytes per token, and CSA2's attention uses a 128-token sliding window.

Taken together, the report says these changes compress the KV cache 4x overall while task-completion quality is maintained. At the same sequence length, DeepSeek-V4.1-Flash needs only about 1/4 of the runtime KV cache storage and about 1/8 of the persistent KV cache storage that DeepSeek-V4-Flash needed, and the report states its overall performance is superior to DeepSeek-V4-Flash's despite the larger parameter count. The author also reads the architecture through a Recursive Transformer lens, as a way of modifying the query while reusing key-value pairs across a recursive pass, pointing to a separate paper, "On the Future Transformer: Loops Are Not What You Need," for background on that framing. The post separately walks through the model's vision branch, including a worked example in which a 1008x1008-pixel image is encoded into 994 tokens after patch merging and tagging, constrained by a 1024-token-per-image budget; the technical report itself states the model supports input resolutions up to approximately 1344x1344 pixels.

Key facts

  • DeepSeek-V4.1-Flash is a 552B-parameter multimodal mixture-of-experts model with a Causal Encoder-Decoder (CED) architecture, supporting up to 1 million tokens of context.
  • The CED design activates only 8B parameters per token during Prefill, using just 20 of the model's 40 layers, and 16B parameters per token during Decode.
  • CSA2 attention compresses KV cache along three axes (a 512-dimensional shared key/value latent per head, merging 2 sequence positions into 1 cache entry, and cross-layer sharing down to 3 Encoder plus 1 Decoder cache copies); combined with FP4 quantization, overall KV cache is compressed 4x while task quality is maintained, per the report.
  • At the same sequence length, runtime KV cache storage drops to about 1/4 and persistent KV cache storage to about 1/8 of DeepSeek-V4-Flash's, with storage growth of about 890 bytes per token for the main KV and indexer combined.
  • The author reports DeepSeek-V4.1 Flash running at nearly 420 tokens/s, and says a person named Cui indicated all DeepSeek-V4 Pro models would be taken offline.

Why it matters

Long-context agent workflows keep growing the context window while tool calls add heavy Prefill computation, and the resulting key-value cache has to be persisted and moved between HBM and external SSD at growing cost. The report treats this as the bottleneck blocking further scaling and answers it architecturally: DeepSeek-V4.1-Flash's Causal Encoder-Decoder design plus the new CSA2 attention mechanism cut KV cache storage 4x overall while, per the report, preserving task-completion quality, and cut runtime and persistent KV storage to about 1/4 and 1/8 of the prior DeepSeek-V4-Flash respectively.

Who it affects

Teams building or running long-horizon agent systems that accumulate large contexts and repeated tool calls stand to gain the most from lower KV cache storage and movement costs. The change also affects anyone currently deploying DeepSeek-V4 Pro models: the author quotes a person referred to only as Cui saying all of those models would be taken offline, apparently in favor of this release.

How to use it

DeepSeek-V4.1-Flash natively supports multimodal input and contexts of up to 1 million tokens. The source gives no pricing, access tier or availability details, so nothing about cost or how to obtain the model can be stated here; what is confirmed is that the older DeepSeek-V4 Pro line is reportedly being retired, per the author's account of Cui's comment, whose role or affiliation is not given.

How solid is it

This is one engineer's outside analysis of DeepSeek's own technical report, the first piece in a planned series, with the author saying more critical content is still to come in a later chapter. The specific figures reproduced here (parameter counts, compression ratios, storage-per-token numbers) come from the report itself, but this piece includes no benchmark or task-accuracy scores to verify the claim that quality is preserved under compression, and it makes no comparison to non-DeepSeek models.

Risks and caveats

No release date for DeepSeek-V4.1-Flash appears anywhere in the source. The claim that model quality holds up under 4x KV cache compression rests on the report's own wording rather than any published benchmark figures in this piece. The identity and role of "Cui," whose remark anchors the framing that this is a bigger deal than a routine update, is not stated, and the more detailed technical scrutiny the author promises for later installments is not yet available.

“supporting input resolutions up to approximately 1344 ×1344 pixels”

— DeepSeek's DeepSeek-V4.1-Flash technical report