---
title: "DeepSeek-V4.1-Flash: Pushing the Limits of KV Cache Compression"
description: "DeepSeek-V4.1-Flash asks how a million-token agent can keep attention memory from overwhelming accelerator memory, persistent storage, and transfer bandwidth. Read it to see how a causal encoder-decoder shares compressed global context across depth, how three CSA2 layer modes divide cache and indexing work, and why bounded replay makes an explicit approximation to avoid persisting local attention state."
canonical_url: "https://fanout.sh/daily/2026-09-13-deepseek-v4-1-flash"
md_url: "https://fanout.sh/daily/2026-09-13-deepseek-v4-1-flash.md"
last_updated: "2026-09-13"
access: "public"
---

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

DeepSeek-V4.1-Flash asks how a million-token agent can keep attention memory from overwhelming accelerator memory, persistent storage, and transfer bandwidth. Read it to see how a causal encoder-decoder shares compressed global context across depth, how three CSA2 layer modes divide cache and indexing work, and why bounded replay makes an explicit approximation to avoid persisting local attention state.

## Paper details

- Authors: DeepSeek-AI

- Venue: Technical Report 2026

- Track: Inference

- Difficulty: Medium

- Reading time: 30 min

- Original paper: https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash/blob/dba1be0a40aa45a94ad051997016db3960a90277/DeepSeek_V41_Tech_Report.pdf

## What you will learn

- How a 20-layer causal encoder produces shared global context for a 20-layer decoder, reducing prompt-side activation from 16B to 8B parameters per token.

- How CSA2 Full, Reindex, and Reuse modes share global KV and sparse selections while preserving layer-local queries and sliding-window state.

- Why 890-byte global KV and one-eighth persistent cache are separate claims, and where approximate bounded replay can fail.

## Continue exploring

[Calculate KV-cache memory in Fanout](https://fanout.sh/labs/kv-cache): Connect the report's cross-layer sharing, cache precision, and context-length claims to the layer-by-layer memory arithmetic of an inference workload.

---
This representation contains public Fanout content only. Protected Pro lessons, account data, billing, checkout, and pricing are not included.

Browse the public content map: https://fanout.sh/sitemap.md