How much VRAM does step-3.5-flash need?

Its published quantizations range from 38.47 GiB to 195.04 GiB in weights alone, before KV cache and runtime overhead.

Can I run step-3.5-flash on a Radeon RX 6500 XT?

Q: Can I run step-3.5-flash on a Radeon RX 6500 XT?

No. No published quantization of step-3.5-flash fits a Radeon RX 6500 XT at any context with q4_0 KV cache.

Not at these settings. No indexed quantization of step-3.5-flash fits Radeon RX 6500 XT at any context we compute, with q4_0 KV. The smallest shipped quantization is 38.47 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	195.04 GiB	196.7	197.1	197.9	199.6	203.0	209.8
Q6_K	150.81 GiB	152.5	152.9	153.7	155.4	158.8	165.5
Q5_K_L	130.88 GiB	132.5	132.9	133.8	135.5	138.9	145.6
Q5_K_M	130.58 GiB	132.2	132.6	133.5	135.2	138.5	145.3
Q5_K_S	126.55 GiB	128.2	128.6	129.5	131.1	134.5	141.3
Q4_1	115.15 GiB	116.8	117.2	118.1	119.7	123.1	129.9
Q4_K_L	112.07 GiB	113.7	114.1	115.0	116.7	120.0	126.8
Q4_K_M	111.70 GiB	113.3	113.8	114.6	116.3	119.7	126.4
Q4_K	110.56 GiB	112.2	112.6	113.5	115.2	118.5	125.3
Q4_K_S	107.70 GiB	109.3	109.8	110.6	112.3	115.7	122.4
Q4_0	105.60 GiB	107.2	107.7	108.5	110.2	113.6	120.3
IQ4_NL	103.97 GiB	105.6	106.0	106.9	108.6	111.9	118.7
IQ4_XS	98.34 GiB	100.0	100.4	101.2	102.9	106.3	113.1
Q3_K_L	87.26 GiB	88.9	89.3	90.2	91.9	95.2	102.0
Q3_K_M	84.04 GiB	85.7	86.1	86.9	88.6	92.0	98.8
IQ3_M	83.98 GiB	85.6	86.0	86.9	88.6	91.9	98.7
Q3_K_S	80.06 GiB	81.7	82.1	83.0	84.7	88.0	94.8
IQ3_XS	75.46 GiB	77.1	77.5	78.4	80.1	83.4	90.2
IQ3_XXS	72.82 GiB	74.5	74.9	75.7	77.4	80.8	87.5
Q2_K_L	65.26 GiB	66.9	67.3	68.2	69.9	73.2	80.0
Q2_K	64.78 GiB	66.4	66.8	67.7	69.4	72.7	79.5
IQ2_M	59.59 GiB	61.2	61.6	62.5	64.2	67.6	74.3
IQ2_S	51.70 GiB	53.3	53.8	54.6	56.3	59.7	66.4
IQ2_XS	51.40 GiB	53.0	53.5	54.3	56.0	59.4	66.1
IQ2_XXS	44.74 GiB	46.4	46.8	47.6	49.3	52.7	59.5
IQ1_M	40.17 GiB	41.8	42.2	43.1	44.8	48.1	54.9
IQ1_S	38.47 GiB	40.1	40.5	41.4	43.1	46.4	53.2

●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 512-token window rather than the full context.