How much VRAM does DeepSeek-V3.2 need?

Its published quantizations range from 138.82 GiB to 1250.09 GiB in weights alone, before KV cache and runtime overhead.

DeepSeek-V3.2 on Apple M3 Pro

Can I run DeepSeek-V3.2 on a Apple M3 Pro?

Q: Can I run DeepSeek-V3.2 on a Apple M3 Pro?

No. No published quantization of DeepSeek-V3.2 fits a Apple M3 Pro at any context with f16 KV cache.

Not at these settings. No indexed quantization of DeepSeek-V3.2 fits Apple M3 Pro at any context we compute, with f16 KV. The smallest shipped quantization is 138.82 GiB in weights alone, against 12.56 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— 18 GB card, 12.56 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	1250.09 GiB	1251.0	1251.3	1251.8	1252.9	1255.0	1259.3
Q8_0	664.30 GiB	665.2	665.5	666.0	667.1	669.2	673.5
Q6_K	513.45 GiB	514.3	514.6	515.2	516.2	518.4	522.7
Q5_K_M	443.46 GiB	444.4	444.6	445.2	446.2	448.4	452.7
Q5_K_S	430.86 GiB	431.8	432.0	432.6	433.6	435.8	440.1
Q4_1	391.87 GiB	392.8	393.0	393.6	394.6	396.8	401.1
Q4_K_M	377.56 GiB	378.5	378.7	379.3	380.3	382.5	386.8
Q4_K_S	354.89 GiB	355.8	356.1	356.6	357.7	359.8	364.1
Q4_0	353.99 GiB	354.9	355.2	355.7	356.8	358.9	363.2
IQ4_NL	353.10 GiB	354.0	354.3	354.8	355.9	358.0	362.3
IQ4_XS	333.74 GiB	334.6	334.9	335.4	336.5	338.7	342.9
Q3_K_M	298.24 GiB	299.1	299.4	299.9	301.0	303.2	307.4
Q3_K_S	270.50 GiB	271.4	271.7	272.2	273.3	275.4	279.7
UD-IQ3_XXS	254.37 GiB	255.3	255.5	256.1	257.1	259.3	263.6
Q2_K_L	228.72 GiB	229.6	229.9	230.4	231.5	233.6	237.9
Q2_K	228.51 GiB	229.4	229.7	230.2	231.3	233.4	237.7
UD-IQ2_M	212.44 GiB	213.3	213.6	214.1	215.2	217.4	221.7
UD-IQ2_XXS	202.22 GiB	203.1	203.4	203.9	205.0	207.1	211.4
UD-IQ1_M	185.62 GiB	186.5	186.8	187.3	188.4	190.5	194.8
UD-IQ1_S	171.48 GiB	172.4	172.6	173.2	174.3	176.4	180.7
IQ2_XXS	162.59 GiB	163.5	163.8	164.3	165.4	167.5	171.8
UD-TQ1_0	150.20 GiB	151.1	151.4	151.9	153.0	155.1	159.4
IQ1_M	138.82 GiB	139.7	140.0	140.5	141.6	143.7	148.0

●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.