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 q4_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 | 201.5 | 201.8 | 202.2 | 203.0 | 204.7 | 208.1 |
| Q8_0 | 106.67 GiB | 107.5 | 107.7 | 108.1 | 108.9 | 110.6 | 114.0 |
| Q6_K_L | 83.13 GiB | 83.9 | 84.1 | 84.6 | 85.4 | 87.1 | 90.5 |
| Q6_K | 82.67 GiB | 83.5 | 83.7 | 84.1 | 84.9 | 86.6 | 90.0 |
| Q5_K_L | 73.87 GiB | 74.7 | 74.9 | 75.3 | 76.1 | 77.8 | 81.2 |
| Q5_K_M | 71.29 GiB | 72.1 | 72.3 | 72.7 | 73.6 | 75.2 | 78.6 |
| Q5_K_S | 69.16 GiB | 69.9 | 70.2 | 70.6 | 71.4 | 73.1 | 76.5 |
| Q4_1 | 64.35 GiB | 65.1 | 65.3 | 65.8 | 66.6 | 68.3 | 71.7 |
| Q4_K_L | 63.62 GiB | 64.4 | 64.6 | 65.0 | 65.9 | 67.6 | 70.9 |
| Q4_K_M | 62.91 GiB | 63.7 | 63.9 | 64.3 | 65.2 | 66.9 | 70.2 |
| Q4_0 | 58.72 GiB | 59.5 | 59.7 | 60.1 | 61.0 | 62.7 | 66.1 |
| IQ4_NL | 58.67 GiB | 59.5 | 59.7 | 60.1 | 60.9 | 62.6 | 66.0 |
| Q4_K_S | 57.23 GiB | 58.0 | 58.2 | 58.7 | 59.5 | 61.2 | 64.6 |
| IQ4_XS | 55.78 GiB | 56.6 | 56.8 | 57.2 | 58.0 | 59.7 | 63.1 |
| Q3_K_L | 53.83 GiB | 54.6 | 54.8 | 55.3 | 56.1 | 57.8 | 61.2 |
| Q3_K_M | 50.59 GiB | 51.4 | 51.6 | 52.0 | 52.9 | 54.5 | 57.9 |
| IQ3_M | 46.87 GiB | 47.7 | 47.9 | 48.3 | 49.1 | 50.8 | 54.2 |
| Q3_K_S | 46.34 GiB | 47.1 | 47.3 | 47.8 | 48.6 | 50.3 | 53.7 |
| IQ3_XS | 44.19 GiB | 45.0 | 45.2 | 45.6 | 46.5 | 48.1 | 51.5 |
| UD-IQ3_XXS | 42.59 GiB | 43.4 | 43.6 | 44.0 | 44.8 | 46.5 | 49.9 |
| IQ3_XXS | 41.87 GiB | 42.7 | 42.9 | 43.3 | 44.1 | 45.8 | 49.2 |
| Q2_K_L | 40.97 GiB | 41.8 | 42.0 | 42.4 | 43.2 | 44.9 | 48.3 |
| Q2_K | 40.03 GiB | 40.8 | 41.0 | 41.5 | 42.3 | 44.0 | 47.4 |
| UD-IQ2_M | 36.39 GiB | 37.2 | 37.4 | 37.8 | 38.7 | 40.3 | 43.7 |
| UD-IQ2_XXS | 34.83 GiB | 35.6 | 35.8 | 36.3 | 37.1 | 38.8 | 42.2 |
| IQ2_M | 34.56 GiB | 35.4 | 35.6 | 36.0 | 36.8 | 38.5 | 41.9 |
| UD-IQ1_M | 32.59 GiB | 33.4 | 33.6 | 34.0 | 34.8 | 36.5 | 39.9 |
| IQ2_S | 31.98 GiB | 32.8 | 33.0 | 33.4 | 34.2 | 35.9 | 39.3 |
| IQ2_XS | 30.68 GiB | 31.5 | 31.7 | 32.1 | 32.9 | 34.6 | 38.0 |
| UD-IQ1_S | 30.24 GiB | 31.0 | 31.2 | 31.7 | 32.5 | 34.2 | 37.6 |
| IQ2_XXS | 28.09 GiB | 28.9 | 29.1 | 29.5 | 30.4 | 32.0 | 35.4 |
| UD-TQ1_0 | 27.25 GiB | 28.0 | 28.3 | 28.7 | 29.5 | 31.2 | 34.6 |
| IQ1_M | 24.51 GiB | 25.3 | 25.5 | 25.9 | 26.8 | 28.5 | 31.8 |
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.