How much VRAM does NuExtract3 need?

Q4_K_M is exactly 2,708,803,456 bytes (2.52 GiB) in weights. Add the KV cache, which depends on context length, plus roughly half a gigabyte of runtime overhead.

How large is NuExtract3's KV cache?

1.00 GiB at 32K context with an f16 cache, computed per layer from the model's architecture. Quantizing the cache to q8_0 roughly halves it.

numind · vision language

NuExtract3

numind/NuExtract3

NuExtract3 at Q4_K_M is exactly 2,708,803,456 bytes (2.52 GiB / 2.71 GB) — an effective 4.774 bits per weight, not the nominal 4. Its KV cache at 32K is 1.00 GiB.

From the file· summed from 1 file(s)From the file· KV per layer

Parameters

4.5B

Architecture

qwen35

32 layers

Context

262,144

native (config.json)

License

apache-2.0

Shipped quantizations

● exact bytes, summed from published files

Quant	Size●	Exact bytes●	Effective bpw●	Tensors●	Publisher
Q2_K	1.78 GiB	1,915,469,696	3.376	—	numind
Q3_K_M	2.11 GiB	2,262,062,976	3.987	—	numind
Q4_K_M	2.52 GiB	2,708,803,456	4.774	—	numind
Q5_K_M	2.86 GiB	3,074,985,856	5.419	—	numind
Q6_K	3.23 GiB	3,464,054,656	6.105	—	numind
Q8_0	4.17 GiB	4,482,402,176	7.900	—	numind

KV cache by context

computed per layer

Context	KV cache (f16)●	Flat formula	Overstated by	Full / windowed / recurrent
4,096	0.13 GiB	0.50 GiB	4.00×	8 / 0 / 24
8,192	0.25 GiB	1.00 GiB	4.00×	8 / 0 / 24
16,384	0.50 GiB	2.00 GiB	4.00×	8 / 0 / 24
32,768	1.00 GiB	4.00 GiB	4.00×	8 / 0 / 24
65,536	2.00 GiB	8.00 GiB	4.00×	8 / 0 / 24
131,072	4.00 GiB	16.00 GiB	4.00×	8 / 0 / 24

24 of 32 layers use linear attention, which keeps a fixed-size recurrent state instead of a per-token cache. Those layers do not grow with context at all — treating them as ordinary attention, as a flat formula does, overstates this model's cache by roughly 4.0× at long context.

Compare with

same modality, comparable size

vs Qwen3.5-9B vs gemma-4-12B-it vs Qwen3.5-4B vs Qwen3.5-0.8B vs gemma-4-E2B-it vs Qwythos-9B-v2 All alternatives →Will it fit on my card? →

Why other calculators give a different number

A parameters × bits ÷ 8 estimate puts Q4_K_M at roughly 2.38 GiB. The real file is 2.52 GiB, because a quantization is a mixture and some tensors are always kept at higher precision.

Architecture

from config.json

Layers

Attention heads

KV heads

Head dim

256

Hidden size

2560

Vocab

248,320

Sliding window

none

SWA period

—

MLA

Experts

—

Experts per token

—

use_sliding_window

—

Questions people ask

How much VRAM does NuExtract3 need?: Q4_K_M is exactly 2,708,803,456 bytes (2.52 GiB) in weights. Add the KV cache, which depends on your context length, plus roughly half a gigabyte of runtime overhead.
How large is NuExtract3's KV cache?: 1.00 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.
Which quantization of NuExtract3 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.