Can I run functionary-medium-v3.2 on a Apple M3 Pro?
Not at these settings. No indexed quantization of functionary-medium-v3.2 fits Apple M3 Pro at any context we compute, with q4_0 KV. The smallest shipped quantization is 15.60 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 | 69.83 GiB | 70.9 | 71.2 | 71.9 | 73.3 | 76.1 | 81.8 |
| Q6_K | 53.91 GiB | 54.9 | 55.3 | 56.0 | 57.4 | 60.2 | 65.8 |
| Q5_K_M | 46.52 GiB | 47.5 | 47.9 | 48.6 | 50.0 | 52.8 | 58.4 |
| Q4_K_L | 40.33 GiB | 41.4 | 41.7 | 42.4 | 43.8 | 46.6 | 52.3 |
| Q4_K_M | 39.60 GiB | 40.6 | 41.0 | 41.7 | 43.1 | 45.9 | 51.5 |
| Q4_K_S | 37.58 GiB | 38.6 | 39.0 | 39.7 | 41.1 | 43.9 | 49.5 |
| IQ4_XS | 35.30 GiB | 36.3 | 36.7 | 37.4 | 38.8 | 41.6 | 47.2 |
| Q3_K_L | 34.59 GiB | 35.6 | 36.0 | 36.7 | 38.1 | 40.9 | 46.5 |
| Q3_K_M | 31.91 GiB | 32.9 | 33.3 | 34.0 | 35.4 | 38.2 | 43.8 |
| IQ3_M | 29.74 GiB | 30.8 | 31.1 | 31.8 | 33.2 | 36.0 | 41.7 |
| Q3_K_S | 28.79 GiB | 29.8 | 30.2 | 30.9 | 32.3 | 35.1 | 40.7 |
| IQ3_XXS | 25.58 GiB | 26.6 | 27.0 | 27.7 | 29.1 | 31.9 | 37.5 |
| Q2_K_L | 25.52 GiB | 26.5 | 26.9 | 27.6 | 29.0 | 31.8 | 37.4 |
| Q2_K | 24.56 GiB | 25.6 | 25.9 | 26.6 | 28.1 | 30.9 | 36.5 |
| IQ2_M | 22.46 GiB | 23.5 | 23.8 | 24.5 | 26.0 | 28.8 | 34.4 |
| IQ2_XS | 19.69 GiB | 20.7 | 21.1 | 21.8 | 23.2 | 26.0 | 31.6 |
| IQ2_XXS | 17.79 GiB | 18.8 | 19.2 | 19.9 | 21.3 | 24.1 | 29.7 |
| IQ1_M | 15.60 GiB | 16.6 | 17.0 | 17.7 | 19.1 | 21.9 | 27.5 |
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.