How much VRAM does Phi-3.5-MoE-instruct need?

Its published quantizations range from 10.27 GiB to 78.00 GiB in weights alone, before KV cache and runtime overhead.

Phi-3.5-MoE-instruct on GeForce RTX 5050

Can I run Phi-3.5-MoE-instruct on a GeForce RTX 5050?

Q: Can I run Phi-3.5-MoE-instruct on a GeForce RTX 5050?

No. No published quantization of Phi-3.5-MoE-instruct fits a GeForce RTX 5050 at any context with q8_0 KV cache.

Not at these settings. No indexed quantization of Phi-3.5-MoE-instruct fits GeForce RTX 5050 at any context we compute, with q8_0 KV. The smallest shipped quantization is 10.27 GiB in weights alone, against 7.44 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— 8 GB card, 7.44 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◐
F16	78.00 GiB	79.1	79.3	79.9	80.9	83.1	87.3
Q8_0	41.44 GiB	42.5	42.8	43.3	44.4	46.5	50.8
Q6_K_L	32.06 GiB	33.1	33.4	33.9	35.0	37.1	41.4
Q6_K	32.00 GiB	33.1	33.3	33.9	34.9	37.1	41.3
Q5_K_L	27.75 GiB	28.8	29.1	29.6	30.7	32.8	37.1
Q5_K_M	27.68 GiB	28.7	29.0	29.5	30.6	32.7	37.0
Q5_K_S	26.84 GiB	27.9	28.2	28.7	29.8	31.9	36.1
Q4_1	24.41 GiB	25.5	25.7	26.3	27.3	29.5	33.7
Q4_K_L	23.70 GiB	24.8	25.0	25.6	26.6	28.8	33.0
Q4_K_M	23.61 GiB	24.7	24.9	25.5	26.5	28.7	32.9
Q4_K_S	22.18 GiB	23.2	23.5	24.0	25.1	27.2	31.5
Q4_0	22.08 GiB	23.1	23.4	23.9	25.0	27.1	31.4
IQ4_NL	21.99 GiB	23.1	23.3	23.9	24.9	27.1	31.3
IQ4_XS	20.78 GiB	21.9	22.1	22.7	23.7	25.8	30.1
Q3_K_L	20.20 GiB	21.3	21.5	22.1	23.1	25.3	29.5
Q3_K_M	18.66 GiB	19.7	20.0	20.5	21.6	23.7	28.0
IQ3_M	17.11 GiB	18.2	18.4	19.0	20.0	22.2	26.4
Q3_K_S	16.82 GiB	17.9	18.2	18.7	19.7	21.9	26.1
IQ3_XS	15.92 GiB	17.0	17.3	17.8	18.9	21.0	25.2
Q2_K_L	14.34 GiB	15.4	15.7	16.2	17.3	19.4	23.6
Q2_K	14.22 GiB	15.3	15.6	16.1	17.1	19.3	23.5
IQ2_M	12.82 GiB	13.9	14.2	14.7	15.7	17.9	22.1
IQ2_S	11.67 GiB	12.7	13.0	13.5	14.6	16.7	21.0
IQ2_XS	11.43 GiB	12.5	12.8	13.3	14.4	16.5	20.7
IQ2_XXS	10.27 GiB	11.3	11.6	12.1	13.2	15.3	19.6

●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 131,072-token window rather than the full context.