How much VRAM does DeepSeek-V4-Flash-0731 need?

UD-IQ1_S is exactly 82,539,237,792 bytes (76.87 GiB) in weights. Add the KV cache, which depends on context length, plus roughly half a gigabyte of runtime overhead.

How large is DeepSeek-V4-Flash-0731's KV cache?

0.06 GiB at 32K context with an f16 cache, computed per layer from the model's architecture — most of its layers use sliding-window attention, so a flat formula overstates this substantially. Quantizing the cache to q8_0 roughly halves it.

Is DeepSeek-V4-Flash-0731 a mixture-of-experts model?

Yes — 256 experts with 6 routed per token. Every expert must be resident in memory, but only the routed ones are read per token, so its memory requirement and its speed behave very differently.

deepseek-ai · text · mixture of experts

DeepSeek-V4-Flash-0731

deepseek-ai/DeepSeek-V4-Flash-0731

DeepSeek-V4-Flash-0731 at UD-IQ1_S is exactly 82,539,237,792 bytes (76.87 GiB / 82.54 GB) — an effective 2.171 bits per weight, not the nominal 1. Its KV cache at 32K is 0.06 GiB, not the 2.69 GiB a flat formula predicts.

From the file· summed from 3 file(s)From the file· KV per layer

Parameters

304B

total, not active

Architecture

deepseek4

43 layers

Context

1,048,576

native (config.json)

License

mit

Shipped quantizations

● exact bytes, summed from published files

Quant	Size●	Exact bytes●	Effective bpw●	Tensors●	Publisher
UD-IQ1_S3 shards	76.87 GiB	82,539,237,792	2.171	—	unsloth
UD-IQ1_M3 shards	80.93 GiB	86,901,313,952	2.285	—	unsloth
UD-IQ2_XXS3 shards	84.62 GiB	90,860,736,928	2.390	—	unsloth
UD-IQ2_M3 shards	84.68 GiB	90,926,928,288	2.391	—	unsloth
UD-IQ3_XXS4 shards	97.05 GiB	104,207,848,032	2.741	—	unsloth
UD-IQ3_S4 shards	108.10 GiB	116,069,339,712	3.053	—	unsloth
UD-Q3_K_M4 shards	119.28 GiB	128,078,484,032	3.369	—	unsloth
UD-IQ4_XS4 shards	127.28 GiB	136,662,446,656	3.594	—	unsloth
UD-IQ4_NL4 shards	127.28 GiB	136,662,446,656	3.594	—	unsloth
MXFP44 shards	145.64 GiB	156,378,345,408	4.113	—	lmstudio-community
MXFP44 shards	145.64 GiB	156,378,347,104	4.113	—	bartowski

KV cache by context

computed per layer — this model uses sliding-window attention

Context	KV cache (f16)●	Flat formula	Overstated by	Full / windowed / recurrent
4,096	0.06 GiB	0.34 GiB	5.33×	0 / 43 / 0
8,192	0.06 GiB	0.67 GiB	10.67×	0 / 43 / 0
16,384	0.06 GiB	1.34 GiB	21.33×	0 / 43 / 0
32,768	0.06 GiB	2.69 GiB	42.67×	0 / 43 / 0
65,536	0.06 GiB	5.38 GiB	85.33×	0 / 43 / 0
131,072	0.06 GiB	10.75 GiB	170.67×	0 / 43 / 0

43 of 43 layers cache only a 128-token window rather than the full context, on a period of 0. Figures assume the default configuration; --swa-full disables the saving entirely.

Compare with

same modality, comparable size

vs Qwen3-Coder-30B-A3B-Instruct vs Qwen3.6-27B vs Qwen3.8-27B vs DeepSeek-V4-Flash vs Qwen3-30B-A3B-Thinking-2507 vs Laguna-XS-2.1 All alternatives →Will it fit on my card? →

Why other calculators give a different number

A parameters × bits ÷ 8 estimate puts UD-IQ1_S at roughly 159.35 GiB. The real file is 76.87 GiB, because a quantization is a mixture and some tensors are always kept at higher precision. The larger discrepancy is the cache: a flat formula gives 2.69 GiB at 32K context where the real figure is 0.06 GiB, because most of this model's layers cache a fixed window rather than the whole context.

Architecture

from config.json

Layers

Attention heads

KV heads

Head dim

512

Hidden size

4096

Vocab

129,280

Sliding window

128

SWA period

MLA

Experts

256

Experts per token

use_sliding_window

—

Questions people ask

How much VRAM does DeepSeek-V4-Flash-0731 need?: UD-IQ1_S is exactly 82,539,237,792 bytes (76.87 GiB) in weights. Add the KV cache, which depends on your context length, plus roughly half a gigabyte of runtime overhead.
How large is DeepSeek-V4-Flash-0731's KV cache?: 0.06 GiB at 32K context with an f16 cache, computed per layer. Quantizing the cache to q8_0 roughly halves it, which is often the difference between a context length fitting and not.
Is DeepSeek-V4-Flash-0731 a mixture-of-experts model?: Yes — 256 experts, 6 routed per token. Every expert must be resident, but only the routed ones are read per token, which is why its memory requirement and its speed behave very differently.
Which quantization of DeepSeek-V4-Flash-0731 should I use?: Q4_K_M is the usual default. Pick the largest quantization that fits your card at the context you actually need — the table above gives exact sizes for every one published.