How much VRAM does DeepSeek-V3.1-Terminus need?

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

DeepSeek-V3.1-Terminus on Radeon RX 6500 XT

Can I run DeepSeek-V3.1-Terminus on a Radeon RX 6500 XT?

Q: Can I run DeepSeek-V3.1-Terminus on a Radeon RX 6500 XT?

No. No published quantization of DeepSeek-V3.1-Terminus fits a Radeon RX 6500 XT at any context with q8_0 KV cache.

Not at these settings. No indexed quantization of DeepSeek-V3.1-Terminus fits Radeon RX 6500 XT at any context we compute, with q8_0 KV. The smallest shipped quantization is 138.82 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◐
F16	1250.09 GiB	1251.2	1251.4	1251.6	1252.2	1253.3	1255.6
BF16	1250.09 GiB	1251.2	1251.4	1251.6	1252.2	1253.3	1255.6
Q8_0	664.30 GiB	665.4	665.6	665.8	666.4	667.6	669.8
Q6_K	514.51 GiB	515.6	515.8	516.1	516.6	517.8	520.0
Q5_K_M	445.49 GiB	446.6	446.7	447.0	447.6	448.7	451.0
Q5_K_S	431.23 GiB	432.4	432.5	432.8	433.3	434.5	436.8
Q4_1	392.12 GiB	393.2	393.4	393.7	394.2	395.4	397.7
Q4_K_M	381.12 GiB	382.2	382.4	382.7	383.2	384.4	386.7
Q4_K_S	367.08 GiB	368.2	368.3	368.6	369.2	370.3	372.6
Q4_0	359.88 GiB	361.0	361.1	361.4	362.0	363.1	365.4
IQ4_NL	354.35 GiB	355.5	355.6	355.9	356.5	357.6	359.9
IQ4_XS	335.26 GiB	336.4	336.5	336.8	337.4	338.5	340.8
Q3_K_M	298.44 GiB	299.6	299.7	300.0	300.6	301.7	304.0
Q3_K_L	297.75 GiB	298.9	299.0	299.3	299.9	301.0	303.3
IQ3_M	286.74 GiB	287.9	288.0	288.3	288.9	290.0	292.3
Q3_K_S	273.21 GiB	274.3	274.5	274.8	275.3	276.5	278.7
IQ3_XS	258.11 GiB	259.2	259.4	259.7	260.2	261.4	263.6
UD-IQ3_XXS	254.71 GiB	255.8	256.0	256.3	256.8	258.0	260.2
IQ3_XXS	249.25 GiB	250.4	250.5	250.8	251.4	252.5	254.8
Q2_K_L	229.03 GiB	230.1	230.3	230.6	231.1	232.3	234.6
Q2_K	228.82 GiB	229.9	230.1	230.4	230.9	232.1	234.4
UD-IQ2_M	213.17 GiB	214.3	214.4	214.7	215.3	216.4	218.7
UD-IQ2_XXS	202.23 GiB	203.4	203.5	203.8	204.3	205.5	207.8
IQ2_M	200.27 GiB	201.4	201.5	201.8	202.4	203.5	205.8
UD-IQ1_M	187.19 GiB	188.3	188.5	188.7	189.3	190.4	192.7
IQ2_XS	181.82 GiB	182.9	183.1	183.4	183.9	185.1	187.4
IQ2_S	176.61 GiB	177.7	177.9	178.2	178.7	179.9	182.1
UD-IQ1_S	173.84 GiB	175.0	175.1	175.4	176.0	177.1	179.4
IQ2_XXS	162.59 GiB	163.7	163.9	164.1	164.7	165.9	168.1
UD-TQ1_0	151.97 GiB	153.1	153.2	153.5	154.1	155.2	157.5
IQ1_M	138.82 GiB	139.9	140.1	140.4	140.9	142.1	144.4

●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.