Can I run Mistral-Small-4-119B-2603 on a Apple M3 Pro?
Not at these settings. No indexed quantization of Mistral-Small-4-119B-2603 fits Apple M3 Pro at any context we compute, with q4_0 KV. The smallest shipped quantization is 25.04 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 | 221.65 GiB | 222.3 | 222.3 | 222.3 | 222.4 | 222.6 | 223.0 |
| Q8_0 | 117.79 GiB | 118.4 | 118.4 | 118.5 | 118.6 | 118.8 | 119.2 |
| Q6_K_L | 95.99 GiB | 96.6 | 96.6 | 96.7 | 96.8 | 97.0 | 97.4 |
| UD-Q6_K | 92.60 GiB | 93.2 | 93.2 | 93.3 | 93.4 | 93.6 | 94.0 |
| Q6_K | 90.96 GiB | 91.6 | 91.6 | 91.6 | 91.7 | 91.9 | 92.3 |
| UD-Q5_K_M | 83.04 GiB | 83.6 | 83.7 | 83.7 | 83.8 | 84.0 | 84.4 |
| Q5_K_L | 79.51 GiB | 80.1 | 80.1 | 80.2 | 80.3 | 80.5 | 80.9 |
| Q5_K_M | 79.20 GiB | 79.8 | 79.8 | 79.9 | 80.0 | 80.2 | 80.6 |
| UD-Q5_K_S | 77.82 GiB | 78.4 | 78.4 | 78.5 | 78.6 | 78.8 | 79.2 |
| Q5_K_S | 76.57 GiB | 77.2 | 77.2 | 77.3 | 77.3 | 77.5 | 77.9 |
| UD-Q4_K_M | 68.70 GiB | 69.3 | 69.3 | 69.4 | 69.5 | 69.7 | 70.1 |
| Q4_K_L | 68.02 GiB | 68.6 | 68.7 | 68.7 | 68.8 | 69.0 | 69.4 |
| Q4_K_M | 67.65 GiB | 68.3 | 68.3 | 68.3 | 68.4 | 68.6 | 69.0 |
| Q4_K_S | 65.08 GiB | 65.7 | 65.7 | 65.8 | 65.9 | 66.1 | 66.5 |
| UD-Q4_K_S | 64.70 GiB | 65.3 | 65.3 | 65.4 | 65.5 | 65.7 | 66.1 |
| Q4_0 | 63.06 GiB | 63.7 | 63.7 | 63.7 | 63.8 | 64.0 | 64.4 |
| IQ4_NL | 62.83 GiB | 63.4 | 63.5 | 63.5 | 63.6 | 63.8 | 64.2 |
| IQ4_XS | 59.41 GiB | 60.0 | 60.0 | 60.1 | 60.2 | 60.4 | 60.8 |
| UD-IQ4_NL | 55.13 GiB | 55.7 | 55.8 | 55.8 | 55.9 | 56.1 | 56.5 |
| UD-IQ4_XS | 54.13 GiB | 54.7 | 54.8 | 54.8 | 54.9 | 55.1 | 55.5 |
| IQ3_M | 53.05 GiB | 53.7 | 53.7 | 53.7 | 53.8 | 54.0 | 54.4 |
| Q3_K_L | 52.99 GiB | 53.6 | 53.6 | 53.7 | 53.8 | 54.0 | 54.4 |
| Q3_K_M | 50.82 GiB | 51.4 | 51.4 | 51.5 | 51.6 | 51.8 | 52.2 |
| UD-Q3_K_M | 50.64 GiB | 51.2 | 51.3 | 51.3 | 51.4 | 51.6 | 52.0 |
| Q3_K_S | 48.42 GiB | 49.0 | 49.1 | 49.1 | 49.2 | 49.4 | 49.8 |
| IQ3_XS | 48.20 GiB | 48.8 | 48.8 | 48.9 | 49.0 | 49.2 | 49.6 |
| IQ3_XXS | 46.44 GiB | 47.0 | 47.1 | 47.1 | 47.2 | 47.4 | 47.8 |
| UD-Q3_K_S | 46.23 GiB | 46.8 | 46.9 | 46.9 | 47.0 | 47.2 | 47.6 |
| UD-IQ3_S | 41.36 GiB | 42.0 | 42.0 | 42.0 | 42.1 | 42.3 | 42.7 |
| UD-IQ3_XXS | 39.86 GiB | 40.5 | 40.5 | 40.5 | 40.6 | 40.8 | 41.2 |
| Q2_K_L | 39.67 GiB | 40.3 | 40.3 | 40.3 | 40.4 | 40.6 | 41.0 |
| Q2_K | 39.18 GiB | 39.8 | 39.8 | 39.9 | 40.0 | 40.2 | 40.6 |
| IQ2_M | 38.77 GiB | 39.4 | 39.4 | 39.5 | 39.5 | 39.7 | 40.1 |
| IQ2_S | 35.52 GiB | 36.1 | 36.2 | 36.2 | 36.3 | 36.5 | 36.9 |
| UD-IQ2_M | 34.99 GiB | 35.6 | 35.6 | 35.7 | 35.8 | 36.0 | 36.4 |
| IQ2_XS | 34.48 GiB | 35.1 | 35.1 | 35.2 | 35.3 | 35.5 | 35.9 |
| UD-IQ2_XXS | 32.53 GiB | 33.1 | 33.2 | 33.2 | 33.3 | 33.5 | 33.9 |
| IQ2_XXS | 31.35 GiB | 32.0 | 32.0 | 32.0 | 32.1 | 32.3 | 32.7 |
| UD-IQ1_M | 30.09 GiB | 30.7 | 30.7 | 30.8 | 30.9 | 31.1 | 31.5 |
| IQ1_M | 27.51 GiB | 28.1 | 28.1 | 28.2 | 28.3 | 28.5 | 28.9 |
| IQ1_S | 25.04 GiB | 25.6 | 25.7 | 25.7 | 25.8 | 26.0 | 26.4 |
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.