How much VRAM does Llama-4-Scout-17B-16E-Instruct-abliterated-v2 need?

Its published quantizations range from 20.66 GiB to 44.25 GiB in weights alone, before KV cache and runtime overhead.

Llama-4-Scout-17B-16E-Instruct-abliterated-v2 on Apple M3 Pro

Can I run Llama-4-Scout-17B-16E-Instruct-abliterated-v2 on a Apple M3 Pro?

Q: Can I run Llama-4-Scout-17B-16E-Instruct-abliterated-v2 on a Apple M3 Pro?

No. No published quantization of Llama-4-Scout-17B-16E-Instruct-abliterated-v2 fits a Apple M3 Pro at any context with q8_0 KV cache.

Not at these settings. No indexed quantization of Llama-4-Scout-17B-16E-Instruct-abliterated-v2 fits Apple M3 Pro at any context we compute, with q8_0 KV. The smallest shipped quantization is 20.66 GiB in weights alone, against 12.56 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— 18 GB card, 12.56 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◐
I1-IQ3_M	44.25 GiB	45.2	45.6	46.4	48.0	51.2	57.6
I1-IQ3_S	43.56 GiB	44.5	44.9	45.7	47.3	50.5	56.9
I1-Q3_K_S	43.53 GiB	44.5	44.9	45.7	47.3	50.5	56.9
Q3_K_S	43.53 GiB	44.5	44.9	45.7	47.3	50.5	56.9
I1-IQ3_XS	41.25 GiB	42.2	42.6	43.4	45.0	48.2	54.6
I1-IQ3_XXS	38.80 GiB	39.8	40.2	41.0	42.6	45.7	52.1
I1-Q2_K	36.85 GiB	37.8	38.2	39.0	40.6	43.8	50.2
Q2_K	36.85 GiB	37.8	38.2	39.0	40.6	43.8	50.2
I1-Q2_K_S	34.42 GiB	35.4	35.8	36.6	38.2	41.4	47.8
I1-IQ2_M	33.04 GiB	34.0	34.4	35.2	36.8	40.0	46.4
I1-IQ2_S	30.07 GiB	31.0	31.4	32.2	33.8	37.0	43.4
I1-IQ2_XS	29.60 GiB	30.6	31.0	31.8	33.4	36.6	42.9
I1-IQ2_XXS	26.60 GiB	27.6	28.0	28.8	30.4	33.5	39.9
I1-IQ1_M	22.88 GiB	23.9	24.3	25.1	26.6	29.8	36.2
I1-IQ1_S	20.66 GiB	21.6	22.0	22.8	24.4	27.6	34.0

●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 8,192-token window rather than the full context.