DeepSeek V4.1 Flash KV Cache: 1.7 GB at FP8 for 1M Tokens, Rivals 43-73 GB

DeepSeek V4.1 Flash needs 1.7 GB of KV cache for a 1M-token context at FP8, Voltgrid calculates from its published config, against 43.4 GB to 72.5 GB for GLM-5.3, Mistral Large 4 and MiniMax M3.1-Flash.

The key-value (KV) cache is the memory a model keeps for every token it has already read, so it does not have to recompute them. In the format DeepSeek actually uses, 4-bit floating point (FP4) with small scale factors, the figure falls to 890 bytes per token, or 0.93 GB for the full context. That is the number on DeepSeek's Hugging Face model card, and The Next Web reported it when the model launched on Sept. 10. Voltgrid's arithmetic from the config file and DeepSeek's reference inference code lands on the same 890 bytes. Every cache figure in this piece is Voltgrid's calculation from published configs, not a measurement on any machine; the weight sizes are the totals of the published checkpoint files, and prices are as reported.

The cache sits in memory on top of the weights, and it grows with every token of context. At 16-bit precision (BF16), GLM-5.3's cache for a 1M-token context comes to 99.9 GB, more than the 96 GB on one NVIDIA RTX PRO 6000 card before a single weight is loaded. DeepSeek's whole cache in its FP4 format would take under 1% of that card. For DeepSeek V4.1 Flash, the 510.3 GB of published weights, not the context, decide which machine can run it. For the other three, a full context adds tens of gigabytes to an already large model.

How Voltgrid Calculated the Cache

For a model with standard grouped-query attention, the cache per token is layers × 2 (one key and one value) × KV heads × head size × bytes per value. GLM-5.3 and Mistral Large 4 use latent attention instead, which stores one compressed 512-value latent plus a 64-value rotary key per layer. All four models also run a sparse-attention indexer, a small side attention that picks which earlier tokens to read, and its 128-value keys are counted as well. Voltgrid multiplied each per-token figure by 1,048,576 tokens, the maximum position count all four configs list, and counts a gigabyte as 10^9 bytes.

BF16 takes 2 bytes per value and 8-bit floating point (FP8) takes 1; the small per-block scale factors that FP8 formats carry are left out. The figures are for one sequence. They exclude fixed-size sliding-window and draft-model caches, which do not grow with the context, and the activations and overhead of the serving software. Weight sizes are the totals of the checkpoint files each maker posted.

Why DeepSeek V4.1 Flash's Cache Is So Small

DeepSeek V4.1 Flash has 40 layers, but its config names only four of them, layers 2, 8, 14 and 20, as sources for the long-range cache, and the layers that look further back than a short window read from those. Three of the four pool every two tokens into one 512-value latent. The fourth, layer 20, keeps one latent per token, because the config's compression ratio changes there from 2 to 1. The same four layers hold the indexer's 128-value keys. That adds up to 1,600 values per token: 3,200 bytes at BF16, or 3.36 GB for 1M tokens, and 1,600 bytes at FP8, or 1.68 GB.

DeepSeek's reference code quantizes the latents to FP4 with one 8-bit scale per 16 values, and the indexer keys to FP4 with one scale per 32 values. Counting the 4-bit values plus those scales gives 890 bytes per token, the figure on DeepSeek's model card. Every layer also keeps a sliding window of the last 128 tokens, a few megabytes in all that does not grow with the context. DeepSeek's model card puts the result at about a quarter of the cache its predecessor, DeepSeek-V4-Flash, needed.

GLM-5.3, Mistral Large 4 and MiniMax M3.1-Flash at 1M Tokens

Z.ai's GLM-5.3 config lists a 512-value latent plus a 64-value rotary key in each of its 78 layers, which a serving engine that keeps the compressed form caches as 576 values. Only 21 layers carry indexer weights; the config marks the other 57 as sharing. That makes 47,616 values per token: 99.9 GB at BF16 and 49.9 GB at FP8 for 1M tokens. Its one multi-token-prediction layer, used only for speculative decoding, would add about 1.5 GB at BF16 when switched on.

Mistral Large 4 has 72 layers, but its config points nine of them at layer 35's cache instead of their own, so 63 layers write a latent. Nine of those 63 skip rotary position encoding and store 512 values rather than 576. With 44 indexers, the total is 41,344 values per token: 86.7 GB at BF16 and 43.4 GB at FP8. Mistral's own parameter file switches on FP8 for both the KV cache and the indexer, so 43.4 GB is the figure closest to how Mistral ships the model.

MiniMax's M3.1-Flash uses grouped-query attention: 60 layers, each with four KV heads of 128 values, holding a key and a value per head. That is 61,440 values per token, or 128.8 GB at BF16 and 64.4 GB at FP8, the largest of the four. Each layer also computes a 128-value index key for MiniMax Sparse Attention. The published files do not say whether that key is kept for every token or once per 128-token block. If it is kept per token, it adds 16.1 GB at BF16 and 8.1 GB at FP8, for up to 145.0 GB and 72.5 GB.

Which Machines Hold the Weights and the Cache

Weights come first. DeepSeek's checkpoint, FP8 with 4-bit experts, totals 510.3 GB of files. Z.ai's FP8 GLM-5.3 totals 755.6 GB, Mistral's FP8 Large 4 totals 1,060.6 GB and Mistral's own NVFP4 version 610.0 GB, and MiniMax's M3.1-Flash, which mixes 8-bit and 4-bit formats, 252.4 GB. With a 1M-token cache, that comes to 511.2 GB for DeepSeek in its FP4 cache format, and at FP8 to 805.6 GB for GLM-5.3, 653.4 GB for Mistral's NVFP4 checkpoint and 316.8 GB to 324.9 GB for MiniMax.

NVIDIA's DGX Station and the GB300 towers built on it, among them HP's ZGX Fury, Dell's Pro Precision with GB300 and MSI's XpertStation WS300, are the only catalog machines with room for three of the four. The Station's 748 GB of coherent memory, 252 GB of HBM3e on the GPU plus 496 GB of LPDDR5X on the Grace CPU in NVIDIA's spec, holds DeepSeek with more than 230 GB to spare, MiniMax with more than 400 GB, and Mistral Large 4 only as the NVFP4 checkpoint. GLM-5.3's FP8 weights alone, 755.6 GB, are already more than the Station's 748 GB, and with a 1M-token cache they need 805.6 GB. None of the four, with its cache, fits in the 252 GB of fast memory, so part of each model sits in the CPU memory at 396 GB/s. StorageReview found that single-stream speed on large mixture-of-experts models stays modest when experts spill there; Voltgrid's DGX Station review has the labs' single-user figures. The lowest US partner price in the catalog, read on Nov. 4, is Exxact's $98,450.

Apple's Mac Studio with the M5 Ultra and 512 GB of unified memory costs at least $22,299, according to MacRumors. On paper it holds MiniMax M3.1-Flash with a full 1M-token cache even at BF16, at 381.3 GB to 397.4 GB, though Apple does not say how much of the 512 GB the GPU may use. DeepSeek's 511.2 GB is nearly the whole 512 GB, too close to call a fit without a measurement, because macOS and the serving software need memory too. Digital Trends reported that the first 512 GB orders arrive on Dec. 3, while AppleInsider and Guru3D reported on Oct. 29 that Apple's store quoted 10 to 12 weeks, with Guru3D citing an estimated delivery of Jan. 28, 2027. MacStories reported a 256K-token GLM-5.3-Flash run failing out of memory on a 256 GB unit, the configuration Voltgrid's Mac Studio review covers.

Smaller machines hold none of the four published checkpoints. A DGX Spark, listed at $6,950, has 128 GB, and GLM-5.3's BF16 cache alone would take most of it. NVIDIA lists the RTX PRO 6000 at $16,000 for 96 GB.

Four RTX PRO 6000 cards give 384 GB, enough on paper for MiniMax's 324.9 GB with an FP8 cache. That four-card setup led the DGX Station only on the earlier MiniMax M3 in Voltgrid's comparison of DGX Station alternatives, and it has not been measured on M3.1-Flash.

Which to Buy

For DeepSeek V4.1 Flash with its full 1M-token context on one machine, a DGX Station-class GB300 tower is the catalog's only clear fit, and the context costs it about 1 GB in DeepSeek's FP4 cache format, or 1.7 GB at FP8, of its 748 GB. Those towers are also the only single boxes in the catalog that hold Mistral Large 4 with a 1M-token FP8 cache, and only in Mistral's NVFP4 checkpoint.

For MiniMax M3.1-Flash at 1M tokens, the 512 GB Mac Studio is the cheapest single box in the catalog with room on paper, at less than a quarter of the Station's lowest partner price, though its 1.2 TB/s of bandwidth trails the Station's 7.1 TB/s of HBM3e. GLM-5.3 fits no single catalog machine at Z.ai's FP8 precision. Buyers of a DGX Spark or one RTX PRO 6000 should not plan on running any of the four published checkpoints with a full context. WiCi's answer for small machines is to stream DeepSeek V4.1 Flash's weights from an SSD, as Voltgrid reported, on the maker's own benchmark.

These figures are for one user. Every extra 1M-token session adds its cache again, so four MiniMax sessions at FP8 would need 257.7 GB of KV cache, or up to 289.9 GB if the index keys are kept per token, on top of the weights, by the same arithmetic. Community quantizations shrink the weights further. The checkpoint and cache figures here are decimal gigabytes of a billion bytes each. The makers do not say whether their memory figures are decimal or binary gigabytes; if binary, each machine has about 7% more room than shown here, so the margins are nominal. Even then, GLM-5.3 with an FP8 cache, at 805.6 GB, would not fit the Station, whose 748 GB would be about 803 decimal gigabytes. The calculation says whether the bytes fit, not how fast a machine reads them at that length.