DeepSeek V4.1 Flash Architecture Breakdown: KV Cache Compressed to 1/437 of First Generation
nashnova research
DeepSeek released V4.1 Flash, scoring 90.6 atop Terminal-Bench 2.1; its 51-page technical report reveals one overriding goal — compressing the KV Cache to 1/437 of the first generation, making long-context agent workloads economically viable.
Why does every innovation point at the KV Cache?
The KV Cache is a model's "memory buffer" during inference — the longer the context, the larger the buffer, and the higher the memory and storage bill. This means → whoever shrinks the buffer furthest can run longer conversations for less money.
From V1 to V4.1 Flash, KV Cache shrank to 1/437 of its original size. Versus the prior generation, high-bandwidth memory (HBM) demand dropped to 1/4 and solid-state drive (SSD) demand to 1/8.
In plain terms = for the same conversation length, the hardware bill is slashed by most of its weight — the step that turns "runs in the lab" into "sells commercially."
Why split input and output processing?
V4.1 Flash uses a Causal Encoder-Decoder (CED) architecture: the first 20 layers only "read" input, the last 20 "write" output. Total parameters: 552 billion; input activation just 8 billion, output activation 16 billion.
The key saving: the decoder's global cache is not computed from each layer's own hidden states — it is projected directly from the encoder's final layer (layer 20). This means → most input tokens only pass through the first 20 layers, not all 40.
For long sequences, prefill computation drops from O(NL) to roughly O(NL/2) — nearly halved. In plain terms = the "prep work" for reading a long document costs about half the compute.
How does the cache compress along three dimensions at once?
V4.1 Flash replaces the prior generation's hybrid attention with CSA2 — Compressed Sparse Attention 2 — compressing along entry size, sequence dimension, and layer dimension simultaneously, reusing global KV and Top-K indices across layers.
CSA2 defines three modes: Full (complete computation, generates its own global KV), Reindex (reuses a prior Full layer's KV but re-scores to pick new Top-K), and Reuse (lightest — directly reuses prior KV and indices).
The decoder adds a hierarchical sparse indexer: the first Full layer builds a shared candidate pool; subsequent layers search only within that pool, so deeper-layer indexing cost no longer grows with context length. This reflects a design philosophy where compression is not a single-point trick but a full-chain rearchitecture of the attention pipeline.
How are cache precision and storage arranged?
The main KV Cache is compressed from FP8 to FP4 (E2M1 scheme), introduced via quantization-aware training (QAT). The more precision-sensitive sliding-window attention KV stays at FP8. In plain terms = where savings are safe, store in "low-res"; where they are not, keep "high-res" — two precision tiers running in parallel.
At deployment, the cache splits into two tiers: sliding-window KV sits in a distributed host-DRAM pool on each machine, with a lifespan of minutes; global KV stays on SSD-backed persistent cache, lasting tens of hours. This means → short-term memory uses fast but expensive RAM, long-term memory uses slow but cheap disk — entirely different cost structures.
What new calls did they make on pre-training and post-training?
Pre-training ran on tens of trillions of tokens of multimodal data with no instability events. DeepSeek made two data-side calls: content from weak models and low-quality machine translations was classified as "implicit duplication" and filtered out; separately, they found their crawlers over-indexed text-only pages and re-steered Common Crawl ingestion toward multimodal sources.
Post-training follows the standard SFT + RL + offline preference distillation (OPD) pipeline, but all improvements concentrate on data and environment tooling. The technical report states: "At the current stage, the marginal returns from engineering data and environment pipelines far exceed those from algorithmic innovation in post-training."
Backing RL scale is DSec, a proprietary sandbox platform running millions of concurrent instances — no Kubernetes, but a custom placement engine trading strong consistency for scalability. This reflects a heavy bet on "infrastructure as competitive advantage."
What does this model mean commercially?
V4.1 Flash is DeepSeek's first production model with native multimodal vision — previously offered only as an experimental branch, now unified into one model.
From September 14, all V4 Pro API requests will be rerouted to V4.1 Flash; until V4.1 Pro launches, this is the only model on the DeepSeek API. This means → DeepSeek is confident enough in this architecture to shift all traffic onto it.
The API exposes three effort tiers controlling reasoning intensity; the distillation stage used over 40 teacher models. In plain terms = whether this "architecture-level cost reduction" can drive large-scale adoption in the developer ecosystem is the key test of its commercial value.
市场有风险，内容仅供研究参考，不构成投资建议。