Can I run Trinity-Large-Preview on a Apple M3 Pro?
Not at these settings. No indexed quantization of Trinity-Large-Preview fits Apple M3 Pro at any context we compute, with q4_0 KV. The smallest shipped quantization is 76.06 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 | 742.60 GiB | 743.4 | 743.5 | 743.7 | 743.9 | 744.5 | 745.5 |
| Q8_0 | 394.59 GiB | 395.4 | 395.5 | 395.6 | 395.9 | 396.4 | 397.5 |
| Q6_K_L | 305.35 GiB | 306.2 | 306.3 | 306.4 | 306.7 | 307.2 | 308.3 |
| Q6_K | 305.07 GiB | 305.9 | 306.0 | 306.1 | 306.4 | 306.9 | 308.0 |
| Q5_K_L | 264.22 GiB | 265.1 | 265.1 | 265.3 | 265.5 | 266.1 | 267.1 |
| Q5_K_M | 263.87 GiB | 264.7 | 264.8 | 264.9 | 265.2 | 265.7 | 266.8 |
| Q5_K_S | 255.88 GiB | 256.7 | 256.8 | 256.9 | 257.2 | 257.7 | 258.8 |
| Q4_1 | 232.76 GiB | 233.6 | 233.7 | 233.8 | 234.1 | 234.6 | 235.7 |
| Q4_K_L | 225.46 GiB | 226.3 | 226.4 | 226.5 | 226.8 | 227.3 | 228.4 |
| Q4_K_M | 225.04 GiB | 225.9 | 226.0 | 226.1 | 226.4 | 226.9 | 227.9 |
| Q4_K_S | 217.12 GiB | 218.0 | 218.1 | 218.2 | 218.4 | 219.0 | 220.0 |
| Q4_0 | 213.32 GiB | 214.2 | 214.2 | 214.4 | 214.6 | 215.2 | 216.2 |
| IQ4_NL | 209.71 GiB | 210.6 | 210.6 | 210.8 | 211.0 | 211.6 | 212.6 |
| IQ4_XS | 198.19 GiB | 199.0 | 199.1 | 199.3 | 199.5 | 200.0 | 201.1 |
| Q3_K_M | 176.43 GiB | 177.3 | 177.4 | 177.5 | 177.8 | 178.3 | 179.3 |
| Q3_K_L | 175.82 GiB | 176.7 | 176.7 | 176.9 | 177.1 | 177.7 | 178.7 |
| UD-IQ3_XXS | 170.01 GiB | 170.8 | 170.9 | 171.1 | 171.3 | 171.9 | 172.9 |
| IQ3_M | 168.61 GiB | 169.5 | 169.5 | 169.7 | 169.9 | 170.5 | 171.5 |
| Q3_K_S | 160.91 GiB | 161.7 | 161.8 | 162.0 | 162.2 | 162.8 | 163.8 |
| IQ3_XS | 151.27 GiB | 152.1 | 152.2 | 152.3 | 152.6 | 153.1 | 154.2 |
| IQ3_XXS | 146.20 GiB | 147.0 | 147.1 | 147.3 | 147.5 | 148.1 | 149.1 |
| UD-IQ2_M | 143.68 GiB | 144.5 | 144.6 | 144.7 | 145.0 | 145.5 | 146.6 |
| UD-IQ2_XXS | 142.94 GiB | 143.8 | 143.9 | 144.0 | 144.3 | 144.8 | 145.9 |
| Q2_K_L | 135.17 GiB | 136.0 | 136.1 | 136.2 | 136.5 | 137.0 | 138.1 |
| Q2_K | 135.04 GiB | 135.9 | 136.0 | 136.1 | 136.4 | 136.9 | 137.9 |
| IQ2_M | 116.50 GiB | 117.3 | 117.4 | 117.6 | 117.8 | 118.4 | 119.4 |
| IQ2_S | 102.49 GiB | 103.3 | 103.4 | 103.6 | 103.8 | 104.3 | 105.4 |
| IQ2_XS | 102.27 GiB | 103.1 | 103.2 | 103.3 | 103.6 | 104.1 | 105.2 |
| IQ2_XXS | 88.58 GiB | 89.4 | 89.5 | 89.6 | 89.9 | 90.4 | 91.5 |
| IQ1_M | 79.29 GiB | 80.1 | 80.2 | 80.4 | 80.6 | 81.1 | 82.2 |
| IQ1_S | 76.06 GiB | 76.9 | 77.0 | 77.1 | 77.4 | 77.9 | 79.0 |
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 4,096-token window rather than the full context.