How much VRAM does GLM-4.5-Air need?

Q4_K_M is exactly 72,975,748,384 bytes (67.96 GiB) in weights. Add the KV cache, which depends on context length, plus roughly half a gigabyte of runtime overhead.

How large is GLM-4.5-Air's KV cache?

5.75 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.

Is GLM-4.5-Air a mixture-of-experts model?

Yes — 128 experts with 8 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.

zai-org · text · mixture of experts

GLM-4.5-Air

zai-org/GLM-4.5-Air

GLM-4.5-Air at Q4_K_M is exactly 72,975,748,384 bytes (67.96 GiB / 72.98 GB) — an effective 5.285 bits per weight, not the nominal 4. Its KV cache at 32K is 5.75 GiB.

From the file· summed from 2 file(s)From the file· KV per layer

Parameters

110B

total, not active

Architecture

glm4moe

46 layers

Context

131,072

native (config.json)

License

mit

Shipped quantizations

● exact bytes, summed from published files

Quant	Size●	Exact bytes●	Effective bpw●	Tensors●	Publisher
IQ1_M	35.26 GiB	37,856,537,056	2.741	—	bartowski
UD-TQ1_0	35.63 GiB	38,260,919,360	2.771	—	unsloth
UD-IQ1_S	35.86 GiB	38,503,140,416	2.788	—	unsloth
IQ2_XXS	36.90 GiB	39,618,734,560	2.869	—	bartowski
UD-IQ1_M	37.31 GiB	40,061,520,960	2.901	—	unsloth
IQ2_XS	39.29 GiB	42,191,153,632	3.055	—	bartowski
IQ2_S	39.62 GiB	42,543,606,240	3.081	—	bartowski
UD-IQ2_XXS	39.75 GiB	42,682,240,064	3.091	—	unsloth
UD-IQ2_M	41.34 GiB	44,389,321,792	3.215	—	unsloth
IQ2_M	42.02 GiB	45,116,025,312	3.267	—	bartowski
Q2_K	42.21 GiB	45,322,103,872	3.282	—	unsloth
Q2_K_L	42.34 GiB	45,467,593,792	3.293	—	unsloth
Q2_K	42.94 GiB	46,102,702,560	3.339	—	bartowski
Q2_K_L	43.50 GiB	46,708,910,560	3.383	—	bartowski
IQ3_XXS2 shards	46.89 GiB	50,343,066,304	3.646	—	bartowski
IQ3_XS2 shards	47.35 GiB	50,839,837,376	3.682	—	bartowski
UD-IQ3_XXS2 shards	47.91 GiB	51,445,288,224	3.726	—	unsloth
Q3_K_S2 shards	49.04 GiB	52,655,381,792	3.813	—	unsloth
Q3_K_S2 shards	49.75 GiB	53,415,959,232	3.868	—	bartowski
IQ3_M2 shards	51.67 GiB	55,477,787,328	4.018	—	bartowski
Q3_K_M2 shards	51.67 GiB	55,481,637,568	4.018	—	bartowski
Q3_K_L2 shards	52.07 GiB	55,910,373,792	4.049	—	lmstudio-community
Q3_K_L2 shards	52.07 GiB	55,910,374,048	4.049	—	bartowski
Q3_K_M2 shards	53.28 GiB	57,203,694,848	4.143	—	unsloth
IQ4_XS2 shards	56.33 GiB	60,480,900,352	4.380	—	unsloth
IQ4_XS2 shards	56.63 GiB	60,806,220,480	4.404	—	bartowski
IQ4_NL2 shards	58.40 GiB	62,710,615,296	4.541	—	unsloth
Q4_02 shards	58.47 GiB	62,781,050,112	4.546	—	unsloth
IQ4_NL2 shards	58.73 GiB	63,064,509,120	4.567	—	bartowski
Q4_02 shards	59.38 GiB	63,761,992,352	4.618	—	bartowski
Q4_K_S2 shards	62.43 GiB	67,033,386,240	4.854	—	unsloth
Q4_K_S2 shards	63.62 GiB	68,310,026,944	4.947	—	bartowski
Q4_12 shards	64.64 GiB	69,407,886,592	5.026	—	unsloth
Q4_12 shards	64.77 GiB	69,546,035,904	5.036	—	bartowski
Q4_K_M2 shards	67.96 GiB	72,975,748,384	5.285	—	unsloth
Q4_K_M2 shards	68.45 GiB	73,501,936,032	5.323	—	lmstudio-community
Q4_K_M2 shards	68.45 GiB	73,501,936,320	5.323	—	bartowski
Q5_K_S2 shards	73.03 GiB	78,417,661,216	5.679	—	unsloth
Q5_K_S2 shards	73.16 GiB	78,553,041,600	5.689	—	bartowski
Q5_K_M2 shards	77.78 GiB	83,519,589,632	6.048	—	unsloth

KV cache by context

computed per layer

Context	KV cache (f16)●	Flat formula	Overstated by	Full / windowed / recurrent
4,096	0.72 GiB	0.72 GiB	—	46 / 0 / 0
8,192	1.44 GiB	1.44 GiB	—	46 / 0 / 0
16,384	2.88 GiB	2.88 GiB	—	46 / 0 / 0
32,768	5.75 GiB	5.75 GiB	—	46 / 0 / 0
65,536	11.50 GiB	11.50 GiB	—	46 / 0 / 0
131,072	23.00 GiB	23.00 GiB	—	46 / 0 / 0

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 gemma-4-E4B-it vs Qwen3-30B-A3B-Thinking-2507 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 57.87 GiB. The real file is 67.96 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

128

Hidden size

4096

Vocab

151,552

Sliding window

none

SWA period

—

MLA

Experts

128

Experts per token

use_sliding_window

—

Questions people ask

How much VRAM does GLM-4.5-Air need?: Q4_K_M is exactly 72,975,748,384 bytes (67.96 GiB) in weights. Add the KV cache, which depends on your context length, plus roughly half a gigabyte of runtime overhead.
How large is GLM-4.5-Air's KV cache?: 5.75 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 GLM-4.5-Air a mixture-of-experts model?: Yes — 128 experts, 8 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 GLM-4.5-Air 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.