Can I run KAT-Dev-72B-Exp on a Apple M3 Pro?
Not at these settings. No indexed quantization of KAT-Dev-72B-Exp fits Apple M3 Pro at any context we compute, with q4_0 KV. The smallest shipped quantization is 22.11 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◐ |
|---|---|---|---|---|---|---|---|
| Q8_0 | 71.96 GiB | 73.0 | 73.3 | 74.0 | 75.4 | 78.3 | 83.9 |
| Q6_K | 59.93 GiB | 61.0 | 61.3 | 62.0 | 63.4 | 66.2 | 71.9 |
| Q5_K_M | 50.71 GiB | 51.7 | 52.1 | 52.8 | 54.2 | 57.0 | 62.6 |
| Q5_K_S | 47.85 GiB | 48.9 | 49.2 | 49.9 | 51.3 | 54.2 | 59.8 |
| Q4_K_L | 45.02 GiB | 46.1 | 46.4 | 47.1 | 48.5 | 51.3 | 57.0 |
| Q4_K_M | 44.16 GiB | 45.2 | 45.5 | 46.2 | 47.7 | 50.5 | 56.1 |
| Q4_1 | 42.56 GiB | 43.6 | 43.9 | 44.6 | 46.1 | 48.9 | 54.5 |
| Q4_K_S | 40.88 GiB | 41.9 | 42.3 | 43.0 | 44.4 | 47.2 | 52.8 |
| Q4_0 | 38.54 GiB | 39.6 | 39.9 | 40.6 | 42.0 | 44.8 | 50.5 |
| IQ4_NL | 38.48 GiB | 39.5 | 39.9 | 40.6 | 42.0 | 44.8 | 50.4 |
| IQ4_XS | 36.98 GiB | 38.0 | 38.4 | 39.1 | 40.5 | 43.3 | 48.9 |
| Q3_K_L | 36.79 GiB | 37.8 | 38.2 | 38.9 | 40.3 | 43.1 | 48.7 |
| Q3_K_M | 35.11 GiB | 36.1 | 36.5 | 37.2 | 38.6 | 41.4 | 47.0 |
| IQ3_M | 33.07 GiB | 34.1 | 34.4 | 35.2 | 36.6 | 39.4 | 45.0 |
| Q3_K_S | 32.12 GiB | 33.2 | 33.5 | 34.2 | 35.6 | 38.4 | 44.0 |
| IQ3_XS | 30.59 GiB | 31.6 | 32.0 | 32.7 | 34.1 | 36.9 | 42.5 |
| IQ3_XXS | 29.66 GiB | 30.7 | 31.0 | 31.7 | 33.2 | 36.0 | 41.6 |
| Q2_K_L | 28.90 GiB | 29.9 | 30.3 | 31.0 | 32.4 | 35.2 | 40.8 |
| Q2_K | 27.76 GiB | 28.8 | 29.1 | 29.9 | 31.3 | 34.1 | 39.7 |
| IQ2_M | 27.32 GiB | 28.4 | 28.7 | 29.4 | 30.8 | 33.6 | 39.3 |
| IQ2_S | 26.02 GiB | 27.1 | 27.4 | 28.1 | 29.5 | 32.3 | 38.0 |
| IQ2_XS | 25.20 GiB | 26.2 | 26.6 | 27.3 | 28.7 | 31.5 | 37.1 |
| IQ2_XXS | 23.74 GiB | 24.8 | 25.1 | 25.8 | 27.2 | 30.0 | 35.7 |
| IQ1_M | 22.11 GiB | 23.1 | 23.5 | 24.2 | 25.6 | 28.4 | 34.0 |
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.