How much VRAM does DeepSeek-Prover-V2-671B need?

Its published quantizations range from 172.10 GiB to 1250.09 GiB in weights alone, before KV cache and runtime overhead.

DeepSeek-Prover-V2-671B on Apple M3 Ultra

Can I run DeepSeek-Prover-V2-671B on a Apple M3 Ultra?

Q: Can I run DeepSeek-Prover-V2-671B on a Apple M3 Ultra?

No. No published quantization of DeepSeek-Prover-V2-671B fits a Apple M3 Ultra at any context with q4_0 KV cache.

Not at these settings. No indexed quantization of DeepSeek-Prover-V2-671B fits Apple M3 Ultra at any context we compute, with q4_0 KV. The smallest shipped quantization is 172.10 GiB in weights alone, against 66.96 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— 96 GB card, 66.96 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	1250.8	1250.9	1251.0	1251.3	1251.9	1253.1
Q8_0	664.30 GiB	665.0	665.1	665.2	665.5	666.1	667.3
Q6_K	513.09 GiB	513.8	513.9	514.0	514.3	514.9	516.1
Q5_K_M	442.78 GiB	443.5	443.6	443.7	444.0	444.6	445.8
Q5_K_S	394.82 GiB	395.5	395.6	395.7	396.1	396.7	397.9
Q4_1	391.10 GiB	391.8	391.9	392.0	392.3	392.9	394.1
Q4_K_M	376.71 GiB	377.4	377.5	377.6	377.9	378.5	379.8
Q4_K_S	353.96 GiB	354.7	354.7	354.9	355.2	355.8	357.0
Q4_0	353.24 GiB	353.9	354.0	354.2	354.5	355.1	356.3
IQ4_NL	352.16 GiB	352.9	352.9	353.1	353.4	354.0	355.2
IQ4_XS	332.72 GiB	333.4	333.5	333.6	334.0	334.6	335.8
Q3_K_M	297.40 GiB	298.1	298.2	298.3	298.6	299.2	300.4
Q3_K_S	269.35 GiB	270.1	270.1	270.3	270.6	271.2	272.4
UD-IQ3_XXS	253.67 GiB	254.4	254.4	254.6	254.9	255.5	256.7
Q2_K_L	227.64 GiB	228.3	228.4	228.6	228.9	229.5	230.7
Q2_K	227.44 GiB	228.1	228.2	228.4	228.7	229.3	230.5
UD-IQ2_M	212.09 GiB	212.8	212.9	213.0	213.3	213.9	215.1
UD-IQ2_XXS	201.13 GiB	201.8	201.9	202.1	202.4	203.0	204.2
UD-IQ1_M	186.06 GiB	186.8	186.8	187.0	187.3	187.9	189.1
UD-IQ1_S	172.10 GiB	172.8	172.9	173.0	173.3	173.9	175.1

●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.