How much VRAM does GLM-5.2 need?

Its published quantizations range from 169.33 GiB to 1404.42 GiB in weights alone, before KV cache and runtime overhead.

GLM-5.2 on Apple M5 Max

Can I run GLM-5.2 on a Apple M5 Max?

Q: Can I run GLM-5.2 on a Apple M5 Max?

No. No published quantization of GLM-5.2 fits a Apple M5 Max at any context with q8_0 KV cache.

Not at these settings. No indexed quantization of GLM-5.2 fits Apple M5 Max at any context we compute, with q8_0 KV. The smallest shipped quantization is 169.33 GiB in weights alone, against 44.64 GiB usable. CPU offload can still run it, slowly.

From the file· weights summed from filesFrom the file· KV computed per layerPredicted· speed and compute buffer

KV cache dtypef16 q8_0 q4_0— 64 GB card, 44.64 GiB usable after overhead

Every quantization at every context

total memory required; green fits, red does not

Quant	Weights●	4K◐	8K◐	16K◐	32K◐	64K◐	128K◐
BF16	1404.42 GiB	1405.2	1405.4	1405.8	1406.5	1407.9	1410.8
Q8_0	746.32 GiB	747.1	747.3	747.6	748.4	749.8	752.7
UD-Q6_K	582.88 GiB	583.7	583.8	584.2	584.9	586.4	589.3
UD-Q5_K_M	522.31 GiB	523.1	523.3	523.6	524.4	525.8	528.7
UD-Q5_K_S	491.05 GiB	491.8	492.0	492.4	493.1	494.6	497.5
UD-Q4_K_M	433.83 GiB	434.6	434.8	435.2	435.9	437.3	440.3
UD-Q4_K_S	406.46 GiB	407.2	407.4	407.8	408.5	410.0	412.9
UD-IQ4_NL	347.07 GiB	347.8	348.0	348.4	349.1	350.6	353.5
UD-IQ4_XS	340.22 GiB	341.0	341.2	341.6	342.3	343.7	346.6
UD-Q3_K_M	319.20 GiB	320.0	320.2	320.5	321.3	322.7	325.6
UD-IQ3_S	287.44 GiB	288.2	288.4	288.8	289.5	291.0	293.9
UD-IQ3_XXS	262.34 GiB	263.1	263.3	263.7	264.4	265.9	268.8
UD-IQ2_M	222.19 GiB	223.0	223.2	223.5	224.2	225.7	228.6
UD-IQ2_XXS	222.08 GiB	222.9	223.0	223.4	224.1	225.6	228.5
UD-IQ1_M	215.35 GiB	216.1	216.3	216.7	217.4	218.9	221.8
UD-IQ1_S	201.83 GiB	202.6	202.8	203.2	203.9	205.3	208.3
Q3_K_M	169.33 GiB	170.1	170.3	170.7	171.4	172.8	175.8

●From the file◐Predictedwhat these mean

Figures are GiB of total memory: weights plus KV cache plus compute buffer and backend overhead. Weights and KV are near-exact; the overhead term is modeled. Hover any cell for the breakdown.

Why other calculators disagree

A parameters × bits ÷ 8 estimate ignores two things that dominate at long context. First, the weights themselves are not the nominal rate — quantizations are mixtures, so the real file is consistently larger than the label implies. Second, this model uses latent attention and allocates no V cache at all, so any formula reading num_key_value_heads overstates its cache by more than an order of magnitude.