Can I run Llama-4-Scout-17B-16E-Instruct on a Apple M3 Pro?
Not at these settings. No indexed quantization of Llama-4-Scout-17B-16E-Instruct fits Apple M3 Pro at any context we compute, with q8_0 KV. The smallest shipped quantization is 24.51 GiB in weights alone, against 12.56 GiB usable. CPU offload can still run it, slowly.
Every quantization at every context
| Quant | Weights● | 4K◐ | 8K◐ | 16K◐ | 32K◐ | 64K◐ | 128K◐ |
|---|---|---|---|---|---|---|---|
| BF16 | 200.76 GiB | 201.7 | 202.1 | 202.9 | 204.5 | 207.7 | 214.1 |
| Q8_0 | 106.67 GiB | 107.6 | 108.0 | 108.8 | 110.4 | 113.6 | 120.0 |
| Q6_K_L | 83.13 GiB | 84.1 | 84.5 | 85.3 | 86.9 | 90.1 | 96.5 |
| Q6_K | 82.67 GiB | 83.6 | 84.0 | 84.8 | 86.4 | 89.6 | 96.0 |
| Q5_K_L | 73.87 GiB | 74.8 | 75.2 | 76.0 | 77.6 | 80.8 | 87.2 |
| Q5_K_M | 71.29 GiB | 72.3 | 72.7 | 73.5 | 75.1 | 78.2 | 84.6 |
| Q5_K_S | 69.16 GiB | 70.1 | 70.5 | 71.3 | 72.9 | 76.1 | 82.5 |
| Q4_1 | 64.35 GiB | 65.3 | 65.7 | 66.5 | 68.1 | 71.3 | 77.7 |
| Q4_K_L | 63.62 GiB | 64.6 | 65.0 | 65.8 | 67.4 | 70.6 | 76.9 |
| Q4_K_M | 62.91 GiB | 63.9 | 64.3 | 65.1 | 66.7 | 69.9 | 76.2 |
| Q4_0 | 58.72 GiB | 59.7 | 60.1 | 60.9 | 62.5 | 65.7 | 72.1 |
| IQ4_NL | 58.67 GiB | 59.6 | 60.0 | 60.8 | 62.4 | 65.6 | 72.0 |
| Q4_K_S | 57.23 GiB | 58.2 | 58.6 | 59.4 | 61.0 | 64.2 | 70.6 |
| IQ4_XS | 55.78 GiB | 56.8 | 57.2 | 57.9 | 59.5 | 62.7 | 69.1 |
| Q3_K_L | 53.83 GiB | 54.8 | 55.2 | 56.0 | 57.6 | 60.8 | 67.2 |
| Q3_K_M | 50.59 GiB | 51.6 | 52.0 | 52.8 | 54.4 | 57.5 | 63.9 |
| IQ3_M | 46.87 GiB | 47.8 | 48.2 | 49.0 | 50.6 | 53.8 | 60.2 |
| Q3_K_S | 46.34 GiB | 47.3 | 47.7 | 48.5 | 50.1 | 53.3 | 59.7 |
| IQ3_XS | 44.19 GiB | 45.2 | 45.6 | 46.4 | 48.0 | 51.1 | 57.5 |
| UD-IQ3_XXS | 42.59 GiB | 43.6 | 44.0 | 44.8 | 46.3 | 49.5 | 55.9 |
| IQ3_XXS | 41.87 GiB | 42.8 | 43.2 | 44.0 | 45.6 | 48.8 | 55.2 |
| Q2_K_L | 40.97 GiB | 41.9 | 42.3 | 43.1 | 44.7 | 47.9 | 54.3 |
| Q2_K | 40.03 GiB | 41.0 | 41.4 | 42.2 | 43.8 | 47.0 | 53.4 |
| UD-IQ2_M | 36.39 GiB | 37.4 | 37.8 | 38.6 | 40.2 | 43.3 | 49.7 |
| UD-IQ2_XXS | 34.83 GiB | 35.8 | 36.2 | 37.0 | 38.6 | 41.8 | 48.2 |
| IQ2_M | 34.56 GiB | 35.5 | 35.9 | 36.7 | 38.3 | 41.5 | 47.9 |
| UD-IQ1_M | 32.59 GiB | 33.6 | 34.0 | 34.8 | 36.3 | 39.5 | 45.9 |
| IQ2_S | 31.98 GiB | 33.0 | 33.4 | 34.1 | 35.7 | 38.9 | 45.3 |
| IQ2_XS | 30.68 GiB | 31.7 | 32.1 | 32.8 | 34.4 | 37.6 | 44.0 |
| UD-IQ1_S | 30.24 GiB | 31.2 | 31.6 | 32.4 | 34.0 | 37.2 | 43.6 |
| IQ2_XXS | 28.09 GiB | 29.1 | 29.5 | 30.3 | 31.9 | 35.0 | 41.4 |
| UD-TQ1_0 | 27.25 GiB | 28.2 | 28.6 | 29.4 | 31.0 | 34.2 | 40.6 |
| IQ1_M | 24.51 GiB | 25.5 | 25.9 | 26.7 | 28.3 | 31.5 | 37.8 |
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.