How much VRAM does Huihui-GLM-4.7-Flash-abliterated-57B need?

Its published quantizations range from 10.79 GiB to 55.03 GiB in weights alone, before KV cache and runtime overhead.

Huihui-GLM-4.7-Flash-abliterated-57B on GeForce RTX 3080

Can I run Huihui-GLM-4.7-Flash-abliterated-57B on a GeForce RTX 3080?

Q: Can I run Huihui-GLM-4.7-Flash-abliterated-57B on a GeForce RTX 3080?

No. No published quantization of Huihui-GLM-4.7-Flash-abliterated-57B fits a GeForce RTX 3080 at any context with q4_0 KV cache.

Not at these settings. No indexed quantization of Huihui-GLM-4.7-Flash-abliterated-57B fits GeForce RTX 3080 at any context we compute, with q4_0 KV. The smallest shipped quantization is 10.79 GiB in weights alone, against 9.30 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— 10 GB card, 9.30 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	55.03 GiB	56.0	56.2	56.5	57.1	58.2	60.6
I1-Q6_K	42.64 GiB	43.6	43.8	44.1	44.7	45.8	48.2
Q6_K	42.64 GiB	43.6	43.8	44.1	44.7	45.8	48.2
I1-Q5_K_M	36.77 GiB	37.8	37.9	38.2	38.8	40.0	42.3
Q5_K_M	36.77 GiB	37.8	37.9	38.2	38.8	40.0	42.3
I1-Q5_K_S	35.72 GiB	36.7	36.9	37.1	37.7	38.9	41.3
Q5_K_S	35.72 GiB	36.7	36.9	37.1	37.7	38.9	41.3
I1-Q4_1	32.47 GiB	33.5	33.6	33.9	34.5	35.7	38.0
I1-Q4_K_M	31.37 GiB	32.4	32.5	32.8	33.4	34.6	36.9
Q4_K_M	31.37 GiB	32.4	32.5	32.8	33.4	34.6	36.9
I1-Q4_K_S	29.54 GiB	30.5	30.7	31.0	31.6	32.7	35.1
Q4_K_S	29.54 GiB	30.5	30.7	31.0	31.6	32.7	35.1
I1-Q4_0	29.37 GiB	30.4	30.5	30.8	31.4	32.6	34.9
IQ4_XS	27.93 GiB	28.9	29.1	29.4	29.9	31.1	33.5
I1-IQ4_XS	27.68 GiB	28.7	28.8	29.1	29.7	30.9	33.2
I1-Q3_K_L	26.94 GiB	27.9	28.1	28.4	29.0	30.1	32.5
Q3_K_L	26.94 GiB	27.9	28.1	28.4	29.0	30.1	32.5
I1-Q3_K_M	24.87 GiB	25.9	26.0	26.3	26.9	28.1	30.4
Q3_K_M	24.87 GiB	25.9	26.0	26.3	26.9	28.1	30.4
I1-IQ3_M	22.85 GiB	23.8	24.0	24.3	24.9	26.0	28.4
I1-IQ3_S	22.52 GiB	23.5	23.7	24.0	24.5	25.7	28.1
I1-Q3_K_S	22.52 GiB	23.5	23.7	24.0	24.5	25.7	28.1
Q3_K_S	22.52 GiB	23.5	23.7	24.0	24.5	25.7	28.1
I1-IQ3_XS	21.34 GiB	22.3	22.5	22.8	23.4	24.5	26.9
I1-IQ3_XXS	20.15 GiB	21.1	21.3	21.6	22.2	23.3	25.7
I1-Q2_K	19.11 GiB	20.1	20.2	20.5	21.1	22.3	24.7
Q2_K	19.11 GiB	20.1	20.2	20.5	21.1	22.3	24.7
I1-Q2_K_S	17.75 GiB	18.7	18.9	19.2	19.8	20.9	23.3
I1-IQ2_M	17.14 GiB	18.1	18.3	18.6	19.2	20.3	22.7
I1-IQ2_S	15.62 GiB	16.6	16.8	17.1	17.6	18.8	21.2
I1-IQ2_XS	15.38 GiB	16.4	16.5	16.8	17.4	18.6	20.9
I1-IQ2_XXS	13.83 GiB	14.8	15.0	15.3	15.8	17.0	19.4
I1-IQ1_M	11.93 GiB	12.9	13.1	13.4	13.9	15.1	17.5
I1-IQ1_S	10.79 GiB	11.8	11.9	12.2	12.8	14.0	16.3

●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, this model uses latent attention and allocates no V cache at all, so any formula reading num_key_value_heads overstates its cache by more than an order of magnitude.