How much VRAM does Gemma-4-Harmonia-31B-uncensored-heretic need?

Its published quantizations range from 11.10 GiB to 57.20 GiB in weights alone, before KV cache and runtime overhead.

Gemma-4-Harmonia-31B-uncensored-heretic on GeForce RTX 5070

Can I run Gemma-4-Harmonia-31B-uncensored-heretic on a GeForce RTX 5070?

Q: Can I run Gemma-4-Harmonia-31B-uncensored-heretic on a GeForce RTX 5070?

No. No published quantization of Gemma-4-Harmonia-31B-uncensored-heretic fits a GeForce RTX 5070 at any context with q8_0 KV cache.

Not at these settings. No indexed quantization of Gemma-4-Harmonia-31B-uncensored-heretic fits GeForce RTX 5070 at any context we compute, with q8_0 KV. The smallest shipped quantization is 11.10 GiB in weights alone, against 11.16 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— 12 GB card, 11.16 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	57.20 GiB	59.0	59.4	60.0	61.4	64.0	69.3
Q8_0	30.39 GiB	32.2	32.6	33.2	34.6	37.2	42.5
Q6_K	23.47 GiB	25.3	25.6	26.3	27.6	30.3	35.6
Q5_K_M	20.35 GiB	22.2	22.5	23.2	24.5	27.2	32.5
Q5_K_S	19.85 GiB	21.7	22.0	22.7	24.0	26.7	32.0
Q4_K_M	17.40 GiB	19.2	19.6	20.2	21.6	24.2	29.5
Q4_K_S	16.54 GiB	18.4	18.7	19.4	20.7	23.4	28.7
IQ4_XS	15.70 GiB	17.5	17.9	18.5	19.9	22.5	27.8
Q3_K_L	15.49 GiB	17.3	17.7	18.3	19.6	22.3	27.6
Q3_K_M	14.24 GiB	16.1	16.4	17.1	18.4	21.1	26.4
Q3_K_S	12.82 GiB	14.7	15.0	15.6	17.0	19.6	24.9
Q2_K	11.10 GiB	12.9	13.3	13.9	15.3	17.9	23.2

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