How much VRAM does AREX-Turbo need?

Q4_K_M is exactly 2,884,850,752 bytes (2.69 GiB) in weights. Add the KV cache, which depends on context length, plus roughly half a gigabyte of runtime overhead.

How large is AREX-Turbo'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.

BAAI · vision language

AREX-Turbo

BAAI/AREX-Turbo

AREX-Turbo at Q4_K_M is exactly 2,884,850,752 bytes (2.69 GiB / 2.88 GB) — an effective 5.084 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
IQ2_M	1.70 GiB	1,825,573,952	3.217	—	bartowski
Q2_K	1.92 GiB	2,063,602,752	3.637	—	bartowski
IQ3_XXS	1.97 GiB	2,115,652,672	3.729	—	bartowski
Q2_K_L	2.07 GiB	2,217,561,152	3.908	—	bartowski
IQ3_XS	2.12 GiB	2,275,775,552	4.011	—	bartowski
Q3_K_S	2.17 GiB	2,326,074,432	4.099	—	bartowski
IQ3_M	2.19 GiB	2,355,156,032	4.151	—	bartowski
Q3_K_M	2.28 GiB	2,447,479,872	4.313	—	bartowski
Q3_K_L	2.36 GiB	2,537,264,192	4.472	—	bartowski
IQ4_XS	2.37 GiB	2,543,162,432	4.482	—	bartowski
IQ4_NL	2.47 GiB	2,647,364,672	4.666	—	bartowski
Q4_0	2.47 GiB	2,647,856,192	4.667	—	bartowski
Q4_K_S	2.53 GiB	2,717,979,712	4.790	—	bartowski
Q4_1	2.66 GiB	2,851,673,152	5.026	—	bartowski
Q4_K_M	2.69 GiB	2,884,850,752	5.084	—	bartowski
Q4_K_L	2.83 GiB	3,038,809,152	5.356	—	bartowski
Q5_K_S	2.91 GiB	3,128,235,072	5.513	—	bartowski
Q5_K_M	3.09 GiB	3,315,586,112	5.843	—	bartowski
Q5_K_L	3.23 GiB	3,469,544,512	6.115	—	bartowski
Q6_K	3.42 GiB	3,677,180,992	6.481	—	bartowski
Q6_K_L	3.57 GiB	3,831,139,392	6.752	—	bartowski
Q8_0	4.19 GiB	4,493,954,112	7.920	—	bartowski
BF16	7.85 GiB	8,424,393,504	14.847	—	bartowski

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 2.69 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 AREX-Turbo need?: Q4_K_M is exactly 2,884,850,752 bytes (2.69 GiB) in weights. Add the KV cache, which depends on your context length, plus roughly half a gigabyte of runtime overhead.
How large is AREX-Turbo'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 AREX-Turbo 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.