How much VRAM does medgemma-27b-text-it need?

Its published quantizations range from 6.06 GiB to 50.32 GiB in weights alone, before KV cache and runtime overhead.

medgemma-27b-text-it on Radeon RX 6500 XT

Can I run medgemma-27b-text-it on a Radeon RX 6500 XT?

Q: Can I run medgemma-27b-text-it on a Radeon RX 6500 XT?

No. No published quantization of medgemma-27b-text-it fits a Radeon RX 6500 XT at any context with q8_0 KV cache.

Not at these settings. No indexed quantization of medgemma-27b-text-it fits Radeon RX 6500 XT at any context we compute, with q8_0 KV. The smallest shipped quantization is 6.06 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	50.32 GiB	51.8	52.0	52.3	52.9	54.3	56.9
Q8_0	26.74 GiB	28.2	28.4	28.7	29.4	30.7	33.4
Q6_K	20.64 GiB	22.1	22.3	22.6	23.3	24.6	27.3
Q5_K_M	17.95 GiB	19.4	19.6	19.9	20.6	21.9	24.6
Q5_K_S	17.48 GiB	18.9	19.1	19.4	20.1	21.4	24.1
Q4_1	15.99 GiB	17.5	17.6	18.0	18.6	19.9	22.6
Q4_K_M	15.41 GiB	16.9	17.0	17.4	18.0	19.4	22.0
Q4_K_S	14.60 GiB	16.1	16.2	16.6	17.2	18.6	21.2
IQ4_NL	14.50 GiB	16.0	16.1	16.5	17.1	18.5	21.1
IQ4_XS	13.75 GiB	15.2	15.4	15.7	16.4	17.7	20.4
Q3_K_L	13.54 GiB	15.0	15.2	15.5	16.2	17.5	20.2
Q3_K_M	12.51 GiB	14.0	14.2	14.5	15.1	16.5	19.1
Q3_K_S	11.33 GiB	12.8	13.0	13.3	14.0	15.3	17.9
UD-IQ3_XXS	10.07 GiB	11.5	11.7	12.0	12.7	14.0	16.7
Q2_K	9.78 GiB	11.3	11.4	11.8	12.4	13.7	16.4
Q2_K_L	9.78 GiB	11.3	11.4	11.8	12.4	13.7	16.4
UD-IQ2_M	8.96 GiB	10.4	10.6	10.9	11.6	12.9	15.6
UD-IQ2_XXS	7.31 GiB	8.8	8.9	9.3	9.9	11.3	13.9
UD-IQ1_M	6.51 GiB	8.0	8.1	8.5	9.1	10.5	13.1
UD-IQ1_S	6.06 GiB	7.5	7.7	8.0	8.7	10.0	12.7

●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.