How much VRAM does Kimi-Linear-48B-A3B-Instruct need?

Its published quantizations range from 9.77 GiB to 91.54 GiB in weights alone, before KV cache and runtime overhead.

Kimi-Linear-48B-A3B-Instruct on Radeon RX 6500 XT

Can I run Kimi-Linear-48B-A3B-Instruct on a Radeon RX 6500 XT?

Q: Can I run Kimi-Linear-48B-A3B-Instruct on a Radeon RX 6500 XT?

No. No published quantization of Kimi-Linear-48B-A3B-Instruct fits a Radeon RX 6500 XT at any context with q8_0 KV cache.

Not at these settings. No indexed quantization of Kimi-Linear-48B-A3B-Instruct fits Radeon RX 6500 XT at any context we compute, with q8_0 KV. The smallest shipped quantization is 9.77 GiB in weights alone, against 3.72 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— 4 GB card, 3.72 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	91.54 GiB	92.5	92.6	92.7	92.9	93.5	94.5
Q8_0	48.66 GiB	49.6	49.7	49.8	50.1	50.6	51.6
Q6_K_L	37.85 GiB	38.8	38.9	39.0	39.3	39.8	40.8
Q6_K	37.68 GiB	38.7	38.7	38.8	39.1	39.6	40.6
Q5_K_L	32.90 GiB	33.9	33.9	34.1	34.3	34.8	35.8
Q5_K_M	32.69 GiB	33.7	33.7	33.8	34.1	34.6	35.6
Q5_K_S	31.68 GiB	32.6	32.7	32.8	33.1	33.6	34.6
Q4_1	28.85 GiB	29.8	29.9	30.0	30.3	30.8	31.8
Q4_K_L	28.26 GiB	29.2	29.3	29.4	29.7	30.2	31.2
Q4_K_M	28.00 GiB	29.0	29.0	29.2	29.4	29.9	30.9
Q4_K_S	26.99 GiB	28.0	28.0	28.1	28.4	28.9	29.9
Q4_0	26.49 GiB	27.5	27.5	27.6	27.9	28.4	29.4
IQ4_NL	26.05 GiB	27.0	27.1	27.2	27.5	28.0	29.0
IQ4_XS	24.65 GiB	25.6	25.7	25.8	26.1	26.6	27.6
Q3_K_L	23.76 GiB	24.7	24.8	24.9	25.2	25.7	26.7
Q3_K_M	21.87 GiB	22.8	22.9	23.0	23.3	23.8	24.8
IQ3_M	21.10 GiB	22.1	22.1	22.3	22.5	23.0	24.0
Q3_K_S	20.12 GiB	21.1	21.2	21.3	21.5	22.0	23.0
IQ3_XS	19.04 GiB	20.0	20.1	20.2	20.4	21.0	22.0
IQ3_XXS	18.30 GiB	19.3	19.3	19.5	19.7	20.2	21.2
Q2_K	16.79 GiB	17.8	17.8	18.0	18.2	18.7	19.7
Q2_K_L	16.67 GiB	17.6	17.7	17.8	18.1	18.6	19.6
IQ2_M	14.71 GiB	15.7	15.7	15.9	16.1	16.6	17.6
IQ2_S	13.01 GiB	14.0	14.0	14.2	14.4	14.9	15.9
IQ2_XS	12.94 GiB	13.9	14.0	14.1	14.4	14.9	15.9
IQ2_XXS	11.30 GiB	12.3	12.3	12.5	12.7	13.2	14.2
IQ1_M	10.17 GiB	11.1	11.2	11.3	11.6	12.1	13.1
IQ1_S	9.77 GiB	10.7	10.8	10.9	11.2	11.7	12.7

●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.