How much VRAM does llm-surgery-dark-arts-gpt-oss-60b-96a12 need?

Its published quantizations range from 31.28 GiB to 58.62 GiB in weights alone, before KV cache and runtime overhead.

llm-surgery-dark-arts-gpt-oss-60b-96a12 on GeForce RTX 5090 D

Can I run llm-surgery-dark-arts-gpt-oss-60b-96a12 on a GeForce RTX 5090 D?

Q: Can I run llm-surgery-dark-arts-gpt-oss-60b-96a12 on a GeForce RTX 5090 D?

No. No published quantization of llm-surgery-dark-arts-gpt-oss-60b-96a12 fits a GeForce RTX 5090 D at any context with q4_0 KV cache.

Not at these settings. No indexed quantization of llm-surgery-dark-arts-gpt-oss-60b-96a12 fits GeForce RTX 5090 D at any context we compute, with q4_0 KV. The smallest shipped quantization is 31.28 GiB in weights alone, against 29.76 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— 32 GB card, 29.76 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	58.62 GiB	59.4	59.5	59.5	59.6	59.8	60.3
I1-Q6_K	58.56 GiB	59.4	59.4	59.5	59.6	59.8	60.2
Q6_K	58.56 GiB	59.4	59.4	59.5	59.6	59.8	60.2
I1-Q5_K_M	44.35 GiB	45.2	45.2	45.2	45.4	45.6	46.0
Q5_K_M	44.35 GiB	45.2	45.2	45.2	45.4	45.6	46.0
I1-Q5_K_S	41.56 GiB	42.4	42.4	42.5	42.6	42.8	43.2
Q5_K_S	41.56 GiB	42.4	42.4	42.5	42.6	42.8	43.2
I1-Q4_K_M	41.48 GiB	42.3	42.3	42.4	42.5	42.7	43.1
Q4_K_M	41.48 GiB	42.3	42.3	42.4	42.5	42.7	43.1
I1-Q4_K_S	38.28 GiB	39.1	39.1	39.2	39.3	39.5	39.9
Q4_K_S	38.28 GiB	39.1	39.1	39.2	39.3	39.5	39.9
I1-Q4_1	34.76 GiB	35.6	35.6	35.7	35.8	36.0	36.4
I1-Q3_K_L	34.73 GiB	35.6	35.6	35.6	35.7	35.9	36.4
Q3_K_L	34.73 GiB	35.6	35.6	35.6	35.7	35.9	36.4
I1-Q3_K_M	33.63 GiB	34.5	34.5	34.5	34.6	34.8	35.3
Q3_K_M	33.63 GiB	34.5	34.5	34.5	34.6	34.8	35.3
IQ4_XS	31.77 GiB	32.6	32.6	32.7	32.8	33.0	33.4
I1-IQ3_M	31.64 GiB	32.5	32.5	32.5	32.6	32.9	33.3
I1-Q2_K_S	31.58 GiB	32.4	32.4	32.5	32.6	32.8	33.2
I1-Q4_0	31.50 GiB	32.3	32.3	32.4	32.5	32.7	33.1
I1-IQ4_XS	31.35 GiB	32.2	32.2	32.3	32.4	32.6	33.0
I1-IQ2_M	31.33 GiB	32.1	32.2	32.2	32.3	32.5	33.0
I1-IQ2_S	31.33 GiB	32.1	32.2	32.2	32.3	32.5	33.0
I1-IQ3_S	31.33 GiB	32.1	32.2	32.2	32.3	32.5	33.0
I1-IQ3_XS	31.33 GiB	32.1	32.2	32.2	32.3	32.5	33.0
I1-IQ3_XXS	31.33 GiB	32.1	32.2	32.2	32.3	32.5	33.0
I1-Q2_K	31.33 GiB	32.1	32.2	32.2	32.3	32.5	33.0
Q2_K	31.33 GiB	32.1	32.2	32.2	32.3	32.5	33.0
I1-Q3_K_S	31.32 GiB	32.1	32.2	32.2	32.3	32.5	33.0
Q3_K_S	31.32 GiB	32.1	32.2	32.2	32.3	32.5	33.0
I1-IQ2_XS	31.29 GiB	32.1	32.1	32.2	32.3	32.5	32.9
I1-IQ1_M	31.28 GiB	32.1	32.1	32.2	32.3	32.5	32.9
I1-IQ1_S	31.28 GiB	32.1	32.1	32.2	32.3	32.5	32.9
I1-IQ2_XXS	31.28 GiB	32.1	32.1	32.2	32.3	32.5	32.9

●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 128-token window rather than the full context.