How much VRAM does cogito-v2-preview-deepseek-671B-MoE need?

Its published quantizations range from 150.73 GiB to 1250.08 GiB in weights alone, before KV cache and runtime overhead.

cogito-v2-preview-deepseek-671B-MoE on Radeon RX 6500 XT

Can I run cogito-v2-preview-deepseek-671B-MoE on a Radeon RX 6500 XT?

Q: Can I run cogito-v2-preview-deepseek-671B-MoE on a Radeon RX 6500 XT?

No. No published quantization of cogito-v2-preview-deepseek-671B-MoE fits a Radeon RX 6500 XT at any context with q4_0 KV cache.

Not at these settings. No indexed quantization of cogito-v2-preview-deepseek-671B-MoE fits Radeon RX 6500 XT at any context we compute, with q4_0 KV. The smallest shipped quantization is 150.73 GiB in weights alone, against 3.72 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— 4 GB card, 3.72 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	1250.08 GiB	1251.1	1251.2	1251.4	1251.7	1252.3	1253.5
Q8_0	664.29 GiB	665.3	665.4	665.6	665.9	666.5	667.7
Q6_K	513.20 GiB	514.3	514.3	514.5	514.8	515.4	516.6
Q5_K_M	443.10 GiB	444.2	444.2	444.4	444.7	445.3	446.5
Q5_K_S	430.45 GiB	431.5	431.6	431.7	432.0	432.6	433.8
Q4_1	391.51 GiB	392.6	392.6	392.8	393.1	393.7	394.9
Q4_K_M	377.13 GiB	378.2	378.3	378.4	378.7	379.3	380.5
Q4_K_S	354.38 GiB	355.4	355.5	355.7	356.0	356.6	357.8
Q4_0	353.48 GiB	354.5	354.6	354.8	355.1	355.7	356.9
IQ4_NL	352.57 GiB	353.6	353.7	353.9	354.2	354.8	356.0
IQ4_XS	333.13 GiB	334.2	334.3	334.4	334.7	335.3	336.5
Q3_K_M	297.88 GiB	298.9	299.0	299.2	299.5	300.1	301.3
Q3_K_S	269.83 GiB	270.9	271.0	271.1	271.4	272.0	273.2
Q2_K_L	228.17 GiB	229.2	229.3	229.4	229.7	230.4	231.6
Q2_K	227.97 GiB	229.0	229.1	229.2	229.5	230.2	231.4
UD-IQ2_M	212.82 GiB	213.9	213.9	214.1	214.4	215.0	216.2
UD-IQ2_XXS	201.88 GiB	202.9	203.0	203.2	203.5	204.1	205.3
UD-IQ1_M	186.85 GiB	187.9	188.0	188.1	188.4	189.0	190.2
UD-IQ1_S	173.12 GiB	174.2	174.3	174.4	174.7	175.3	176.5
UD-TQ1_0	150.73 GiB	151.8	151.9	152.0	152.3	152.9	154.1

●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.