DeepSeek unveils V4.1-Flash compression design, reducing global KV cache to 890 bytes per token

DeepSeek released its V4.1-Flash technical report on September 17, detailing how long-horizon agents can reduce the cost of storing and rereading context. The model supports one million tokens, and the report divides its optimizations into three areas—prefill computation, global cache, and persistent state—so deployers can assess compute, GPU memory, and disk bottlenecks separately. [Technical report](https://arxiv.org/abs/2609.19969)

Its causal encoder–decoder architecture splits the 40-layer network into two halves of 20 layers each. The decoder’s global keys and values are produced by projecting the encoder’s final states, allowing most inputs to skip the full computation through the second half. Active parameters per token are therefore 8B during prefill and 16B during decoding. However, the backbone still contains 552B parameters, plus Engram conditional memory, so hardware requirements cannot be assessed from active parameter counts alone. [Model card](https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash)

For caching, CSA2 lets different layers choose whether to construct their own keys and values, share keys and values but perform indexing again, or also reuse the sparse attention selection results. Combined with an FP4 main cache, this reduces the global KV cache to 890 bytes per token, roughly one-quarter of the previous generation’s size. This figure covers only the global cache; deployment sizing must also account for local state, weights, and other runtime memory. [Architecture overview](https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash)

Another design replays only the most recent sliding window to approximately reconstruct local attention states, eliminating the need to retain those states on SSD. The authors say this reduces persistent KV storage to roughly one-eighth of the previous generation’s footprint, but the reconstructed states are not mathematically identical to the originals. The paper also acknowledges that sparse selection and cache restoration boundaries may still cause capability degradation that testing has yet to detect. [Reconstruction method and limitations](https://arxiv.org/html/2609.19969v1)

The publicly available minimal inference implementation helps with inspecting the architecture, but it still uses standard autoregressive generation. Even though it includes a forward path for a draft model, this does not establish that it reproduces the full speculative decoding performance. Its built-in self-tests use uninitialized weights and check only tensor shapes and kernel integration, not numerical correctness. [Inference implementation](https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash/blob/main/inference/README.md)

From an engineering perspective, this design trades storage costs for a limited amount of recomputation and is particularly worth testing with tool responses across multiple turns. The next step is to compare latency on cache hits and during restoration, sparse retrieval accuracy over long texts, and total GPU memory usage before determining whether compression translates into greater serving capacity. For agents whose inputs keep growing, prefix retention, transfer across nodes, and restoration quality must be evaluated together. Generation speed measured with short prompts cannot capture the full cost of reusing historical state in long-running tasks.