How much VRAM does Qwen2.5-Math-72B-Instruct need?

Its published quantizations range from 22.11 GiB to 71.96 GiB in weights alone, before KV cache and runtime overhead.

Qwen2.5-Math-72B-Instruct on Radeon RX 6500 XT

Can I run Qwen2.5-Math-72B-Instruct on a Radeon RX 6500 XT?

Q: Can I run Qwen2.5-Math-72B-Instruct on a Radeon RX 6500 XT?

No. No published quantization of Qwen2.5-Math-72B-Instruct fits a Radeon RX 6500 XT at any context with q8_0 KV cache.

Not at these settings. No indexed quantization of Qwen2.5-Math-72B-Instruct fits Radeon RX 6500 XT at any context we compute, with q8_0 KV. The smallest shipped quantization is 22.11 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◐
Q8_0	71.96 GiB	73.7	74.3	75.6	78.3	83.6	94.2
Q6_K	59.93 GiB	61.6	62.3	63.6	66.3	71.6	82.2
Q5_K_M	50.71 GiB	52.4	53.1	54.4	57.1	62.4	73.0
Q4_K_M	44.16 GiB	45.9	46.5	47.8	50.5	55.8	66.4
Q4_0	38.54 GiB	40.2	40.9	42.2	44.9	50.2	60.8
IQ4_XS	36.98 GiB	38.7	39.3	40.7	43.3	48.6	59.3
Q3_K_L	36.79 GiB	38.5	39.2	40.5	43.1	48.4	59.1
Q3_K_M	35.11 GiB	36.8	37.5	38.8	41.5	46.8	57.4
IQ3_M	33.07 GiB	34.8	35.4	36.8	39.4	44.7	55.3
Q3_K_S	32.12 GiB	33.8	34.5	35.8	38.5	43.8	54.4
IQ3_XXS	29.66 GiB	31.4	32.0	33.3	36.0	41.3	51.9
Q2_K_L	28.90 GiB	30.6	31.3	32.6	35.2	40.6	51.2
Q2_K	27.76 GiB	29.5	30.1	31.5	34.1	39.4	50.0
IQ2_M	27.32 GiB	29.0	29.7	31.0	33.7	39.0	49.6
IQ2_XS	25.20 GiB	26.9	27.6	28.9	31.5	36.9	47.5
IQ2_XXS	23.74 GiB	25.4	26.1	27.4	30.1	35.4	46.0
IQ1_M	22.11 GiB	23.8	24.5	25.8	28.5	33.8	44.4

●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, most of this model's layers cache only a 4,096-token window rather than the full context.