How much VRAM does GLM-4.7-REAP-218B-A32B need?

Q4_K_M is exactly 132,044,138,240 bytes (122.98 GiB) in weights. Add the KV cache, which depends on context length, plus roughly half a gigabyte of runtime overhead.

How large is GLM-4.7-REAP-218B-A32B's KV cache?

11.50 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-4.7-REAP-218B-A32B a mixture-of-experts model?

Yes — 96 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.

cerebras · text · mixture of experts

GLM-4.7-REAP-218B-A32B

cerebras/GLM-4.7-REAP-218B-A32B

GLM-4.7-REAP-218B-A32B at Q4_K_M is exactly 132,044,138,240 bytes (122.98 GiB / 132.04 GB) — an effective 4.837 bits per weight, not the nominal 4. Its KV cache at 32K is 11.50 GiB.

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

Parameters

218B

total, not active

Architecture

glm4moe

92 layers

Context

202,752

native (config.json)

License

mit

Shipped quantizations

● exact bytes, summed from published files

Quant	Size●	Exact bytes●	Effective bpw●	Tensors●	Publisher
IQ1_S	44.60 GiB	47,890,056,416	1.754	—	bartowski
IQ1_M2 shards	46.56 GiB	49,996,465,568	1.831	—	bartowski
UD-TQ1_0	49.34 GiB	52,974,155,200	1.941	—	unsloth
IQ2_XXS2 shards	51.57 GiB	55,376,889,280	2.029	—	bartowski
UD-IQ1_S2 shards	56.46 GiB	60,618,274,432	2.221	—	unsloth
IQ2_XS2 shards	58.74 GiB	63,075,075,488	2.311	—	bartowski
IQ2_S2 shards	59.49 GiB	63,877,399,968	2.340	—	bartowski
UD-IQ1_M2 shards	62.63 GiB	67,251,951,232	2.464	—	unsloth
IQ2_M2 shards	67.01 GiB	71,947,175,328	2.636	—	bartowski
UD-IQ2_XXS2 shards	67.13 GiB	72,082,814,592	2.641	—	unsloth
UD-IQ2_M2 shards	70.78 GiB	75,999,245,952	2.784	—	unsloth
Q2_K2 shards	73.81 GiB	79,257,367,968	2.903	—	bartowski
Q2_K_L3 shards	74.52 GiB	80,015,128,128	2.931	—	bartowski
Q2_K2 shards	74.91 GiB	80,434,292,352	2.946	—	unsloth
Q2_K_L2 shards	75.08 GiB	80,616,154,752	2.953	—	unsloth
IQ3_XXS3 shards	82.27 GiB	88,333,121,024	3.236	—	bartowski
UD-IQ3_XXS2 shards	83.34 GiB	89,486,411,392	3.278	—	unsloth
IQ3_XS3 shards	85.49 GiB	91,797,190,208	3.363	—	bartowski
Q3_K_S2 shards	88.26 GiB	94,772,749,952	3.472	—	unsloth
Q3_K_S3 shards	90.39 GiB	97,050,392,096	3.555	—	bartowski
IQ3_M3 shards	94.52 GiB	101,489,473,056	3.718	—	bartowski
Q3_K_M3 shards	94.69 GiB	101,676,578,336	3.725	—	bartowski
Q3_K_L3 shards	97.31 GiB	104,490,038,816	3.828	—	bartowski
Q3_K_M3 shards	97.57 GiB	104,768,300,800	3.838	—	unsloth
IQ4_XS3 shards	108.90 GiB	116,930,144,000	4.284	—	unsloth
IQ4_XS3 shards	110.09 GiB	118,209,508,928	4.330	—	bartowski
IQ4_NL3 shards	115.13 GiB	123,623,991,040	4.529	—	unsloth
Q4_03 shards	115.44 GiB	123,952,981,760	4.541	—	unsloth
Q4_K_S3 shards	115.81 GiB	124,344,887,008	4.555	—	unsloth
IQ4_NL4 shards	116.20 GiB	124,764,092,032	4.571	—	bartowski
Q4_04 shards	117.58 GiB	126,245,205,664	4.625	—	bartowski
Q4_K_S4 shards	120.24 GiB	129,106,507,392	4.729	—	bartowski
Q4_K_M3 shards	122.98 GiB	132,044,138,240	4.837	—	unsloth
Q4_K_M4 shards	124.65 GiB	133,841,155,776	4.903	—	bartowski
Q4_13 shards	127.62 GiB	137,030,035,200	5.020	—	unsloth
Q4_14 shards	128.16 GiB	137,611,114,112	5.041	—	bartowski
Q5_K_S4 shards	140.33 GiB	150,682,290,048	5.520	—	unsloth
Q5_K_S4 shards	140.74 GiB	151,120,295,616	5.536	—	bartowski
Q5_K_M4 shards	144.40 GiB	155,051,001,728	5.680	—	unsloth
Q5_K_M4 shards	145.23 GiB	155,940,959,936	5.713	—	bartowski

KV cache by context

computed per layer

Context	KV cache (f16)●	Flat formula	Overstated by	Full / windowed / recurrent
4,096	1.44 GiB	1.44 GiB	—	92 / 0 / 0
8,192	2.88 GiB	2.88 GiB	—	92 / 0 / 0
16,384	5.75 GiB	5.75 GiB	—	92 / 0 / 0
32,768	11.50 GiB	11.50 GiB	—	92 / 0 / 0
65,536	23.00 GiB	23.00 GiB	—	92 / 0 / 0
131,072	46.00 GiB	46.00 GiB	—	92 / 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 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 Q4_K_M at roughly 114.40 GiB. The real file is 122.98 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

5120

Vocab

151,552

Sliding window

none

SWA period

—

MLA

Experts

Experts per token

use_sliding_window

—

Questions people ask

How much VRAM does GLM-4.7-REAP-218B-A32B need?: Q4_K_M is exactly 132,044,138,240 bytes (122.98 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-4.7-REAP-218B-A32B's KV cache?: 11.50 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-4.7-REAP-218B-A32B a mixture-of-experts model?: Yes — 96 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-4.7-REAP-218B-A32B 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.