How much VRAM does DeepSeek-Coder-V2-Instruct need?

Its published quantizations range from 44.14 GiB to 439.20 GiB in weights alone, before KV cache and runtime overhead.

DeepSeek-Coder-V2-Instruct on Apple M5 Pro

Can I run DeepSeek-Coder-V2-Instruct on a Apple M5 Pro?

Q: Can I run DeepSeek-Coder-V2-Instruct on a Apple M5 Pro?

No. No published quantization of DeepSeek-Coder-V2-Instruct fits a Apple M5 Pro at any context with q4_0 KV cache.

Not at these settings. No indexed quantization of DeepSeek-Coder-V2-Instruct fits Apple M5 Pro at any context we compute, with q4_0 KV. The smallest shipped quantization is 44.14 GiB in weights alone, against 33.48 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— 48 GB card, 33.48 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	439.20 GiB	439.9	439.9	440.1	440.4	441.0	442.2
Q8_0	233.41 GiB	234.1	234.1	234.3	234.6	235.2	236.4
Q6_K	180.25 GiB	180.9	181.0	181.1	181.4	182.0	183.2
Q5_K_M	155.74 GiB	156.4	156.5	156.6	156.9	157.5	158.7
Q5_K	155.74 GiB	156.4	156.5	156.6	156.9	157.5	158.7
Q5_0	151.16 GiB	151.8	151.9	152.0	152.3	152.9	154.1
Q5_K_S	151.16 GiB	151.8	151.9	152.0	152.3	152.9	154.1
Q4_K	132.67 GiB	133.3	133.4	133.6	133.9	134.4	135.6
Q4_K_M	132.67 GiB	133.3	133.4	133.6	133.9	134.4	135.6
Q4_K_S	124.68 GiB	125.3	125.4	125.6	125.9	126.5	127.6
IQ4_NL	123.78 GiB	124.4	124.5	124.7	125.0	125.6	126.7
Q4_0	123.78 GiB	124.4	124.5	124.7	125.0	125.6	126.7
IQ4_XS	116.94 GiB	117.6	117.7	117.8	118.1	118.7	119.9
Q3_K_L	113.97 GiB	114.6	114.7	114.9	115.2	115.7	116.9
Q3_K	104.93 GiB	105.6	105.7	105.8	106.1	106.7	107.9
Q3_K_M	104.93 GiB	105.6	105.7	105.8	106.1	106.7	107.9
IQ3_M	96.27 GiB	96.9	97.0	97.2	97.5	98.0	99.2
IQ3_S	94.70 GiB	95.4	95.4	95.6	95.9	96.5	97.7
Q3_K_S	94.70 GiB	95.4	95.4	95.6	95.9	96.5	97.7
IQ3_XS	89.69 GiB	90.4	90.4	90.6	90.9	91.5	92.7
IQ3_XXS	84.61 GiB	85.3	85.3	85.5	85.8	86.4	87.6
Q2_K_L	81.44 GiB	82.1	82.2	82.3	82.6	83.2	84.4
Q2_K	80.04 GiB	80.7	80.8	80.9	81.2	81.8	83.0
Q2_K_S	74.13 GiB	74.8	74.9	75.0	75.3	75.9	77.1
IQ2_M	71.64 GiB	72.3	72.4	72.5	72.8	73.4	74.6
IQ2_S	65.07 GiB	65.7	65.8	66.0	66.3	66.8	68.0
IQ2_XS	63.99 GiB	64.7	64.7	64.9	65.2	65.8	67.0
IQ2_XXS	57.28 GiB	57.9	58.0	58.2	58.5	59.1	60.2
IQ1_M	49.06 GiB	49.7	49.8	50.0	50.2	50.8	52.0
IQ1_S	44.14 GiB	44.8	44.9	45.0	45.3	45.9	47.1

●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.