I Tested Deepseek V4.1's Cache Innovation

DeepSeek shrank the KV cache 437 times, and the model they divided by is from 2023. On September 10th DeepSeek released V4.1-Flash, a 763 billion parameter multimodal MoE that holds a million tokens of context at 890 bytes a token. The technical report calls that 437 times smaller than DeepSeek V1, and the figure reproduces exactly from two public config files: V1 keeps 95 layers of 8 key value heads at 128 wide in bfloat16, which is 389,120 bytes a token, and 389,120 divided by 890 is 437.2. What the coverage leaves out is what V1 was. It shipped in November 2023 with a 4,096 token context window, before multi-head latent attention existed. Against the model V4.1-Flash actually replaces, the same paper reports four times, not 437. Fill both caches completely and the honest comparison is better anyway: V1 full is 1.59 GB, V4.1-Flash full is 0.93 GB, which is 256 times the context in 41% less memory. The comparison that decides your bill is a different one. Divide 890 bytes into the weights on the same card. V4.1-Flash is 510.3 GB across 48 files, with roughly 313 GB resident once the 196.9 billion parameter engram lookup tables move to host memory. On eight NVIDIA H200 SXM cards at 141 GB each, that leaves 815 GB free, or about 915 concurrent users each holding a full million tokens. DeepSeek V3 on the same node serves 40. That is where three tenths of a cent per million cached input tokens comes from, and it is also why the same model produces one token every 23 seconds on a 16 GB Mac Mini. Timestamps 00:00 — The 437× KV Cache Claim 01:46 — How DeepSeek Gets to 437× Smaller 03:34 — The Four Architectural Tricks 06:03 — 256× More Context With Less Memory 08:23 — 915 Users vs. 40: The Real Impact 09:25 — Why DeepSeek Flash Is Terrible for Solo Developers 11:00 — The Hidden Tradeoffs and What Actually Matters WHAT THIS VIDEO COVERS The 437x figure reproduces exactly, and its denominator is a three-year-old 4,096 token model Only four of V4.1-Flash's 40 layers own a KV cache: layers 2, 8, 14 and 20 On eight H200s the weights are 352 times one user's full million-token cache Concurrency, not compression, is what turns 890 bytes into three tenths of a cent DeepSeek retired V4-Flash, not V4-Pro, and reversed the V4-Pro plan before the cutover The paper's own limitations section names two failure modes nobody quotes SOURCES Technical report: https://arxiv.org/abs/2609.19969  Model and config.json: https://huggingface.co/deepseek-ai/De... DeepSeek V1 config.json: https://huggingface.co/deepseek-ai/de... DeepSeek V3 config.json: https://huggingface.co/deepseek-ai/De... Official announcement: https://deepseek.com/en/news/deepseek... Live API pricing: https://api-docs.deepseek.com/quick_s... Parameter breakdown: https://www.latent.space/p/ainews-dee... Local hardware reality check: https://www.modemguides.com/blogs/ai-... Throughput and verbosity: https://artificialanalysis.ai/models/... #deepseek #deepselect #cuda #gpu #pytorch #llmarchitecture #kvcache #nvidia #blackwell #kaiexplainsyt