How much VRAM does gemma-4-A4B-98e-v7-coder-it need?

Its published quantizations range from 8.22 GiB to 37.06 GiB in weights alone, before KV cache and runtime overhead.

gemma-4-A4B-98e-v7-coder-it on Radeon RX 6500 XT

Can I run gemma-4-A4B-98e-v7-coder-it on a Radeon RX 6500 XT?

Q: Can I run gemma-4-A4B-98e-v7-coder-it on a Radeon RX 6500 XT?

No. No published quantization of gemma-4-A4B-98e-v7-coder-it fits a Radeon RX 6500 XT at any context with q8_0 KV cache.

Not at these settings. No indexed quantization of gemma-4-A4B-98e-v7-coder-it fits Radeon RX 6500 XT at any context we compute, with q8_0 KV. The smallest shipped quantization is 8.22 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	37.06 GiB	38.2	38.3	38.4	38.8	39.4	40.8
Q8_0	19.71 GiB	20.8	20.9	21.1	21.4	22.1	23.4
Q6_K_L	16.75 GiB	17.9	18.0	18.1	18.5	19.1	20.4
Q6_K	16.58 GiB	17.7	17.8	18.0	18.3	19.0	20.3
Q5_K_L	14.20 GiB	15.3	15.4	15.6	15.9	16.6	17.9
Q5_K_M	14.04 GiB	15.2	15.2	15.4	15.7	16.4	17.7
Q4_K_L	12.50 GiB	13.6	13.7	13.9	14.2	14.9	16.2
Q4_K_M	12.33 GiB	13.5	13.5	13.7	14.0	14.7	16.0
Q4_K_S	11.37 GiB	12.5	12.6	12.7	13.1	13.7	15.1
IQ4_NL	10.63 GiB	11.8	11.8	12.0	12.3	13.0	14.3
IQ4_XS	10.25 GiB	11.4	11.5	11.6	12.0	12.6	14.0
Q3_K_L	10.19 GiB	11.3	11.4	11.6	11.9	12.6	13.9
Q3_K_M	9.79 GiB	10.9	11.0	11.2	11.5	12.2	13.5
Q2_K	8.22 GiB	9.3	9.4	9.6	9.9	10.6	11.9

●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 1,024-token window rather than the full context.