Can I run Trinity-Large-Preview on a Apple M5 Max?
Not at these settings. No indexed quantization of Trinity-Large-Preview fits Apple M5 Max at any context we compute, with q8_0 KV. The smallest shipped quantization is 76.06 GiB in weights alone, against 25.11 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 | 743.7 | 743.8 | 744.1 | 744.6 | 745.6 | 747.6 |
| Q8_0 | 394.59 GiB | 395.7 | 395.8 | 396.1 | 396.6 | 397.6 | 399.6 |
| Q6_K_L | 305.35 GiB | 306.4 | 306.6 | 306.8 | 307.3 | 308.3 | 310.3 |
| Q6_K | 305.07 GiB | 306.1 | 306.3 | 306.6 | 307.1 | 308.1 | 310.1 |
| Q5_K_L | 264.22 GiB | 265.3 | 265.5 | 265.7 | 266.2 | 267.2 | 269.2 |
| Q5_K_M | 263.87 GiB | 264.9 | 265.1 | 265.4 | 265.9 | 266.9 | 268.8 |
| Q5_K_S | 255.88 GiB | 257.0 | 257.1 | 257.4 | 257.9 | 258.9 | 260.9 |
| Q4_1 | 232.76 GiB | 233.8 | 234.0 | 234.3 | 234.8 | 235.7 | 237.7 |
| Q4_K_L | 225.46 GiB | 226.5 | 226.7 | 227.0 | 227.5 | 228.5 | 230.4 |
| Q4_K_M | 225.04 GiB | 226.1 | 226.3 | 226.5 | 227.0 | 228.0 | 230.0 |
| Q4_K_S | 217.12 GiB | 218.2 | 218.4 | 218.6 | 219.1 | 220.1 | 222.1 |
| Q4_0 | 213.32 GiB | 214.4 | 214.6 | 214.8 | 215.3 | 216.3 | 218.3 |
| IQ4_NL | 209.71 GiB | 210.8 | 211.0 | 211.2 | 211.7 | 212.7 | 214.7 |
| IQ4_XS | 198.19 GiB | 199.3 | 199.4 | 199.7 | 200.2 | 201.2 | 203.2 |
| Q3_K_M | 176.43 GiB | 177.5 | 177.7 | 177.9 | 178.4 | 179.4 | 181.4 |
| Q3_K_L | 175.82 GiB | 176.9 | 177.1 | 177.3 | 177.8 | 178.8 | 180.8 |
| UD-IQ3_XXS | 170.01 GiB | 171.1 | 171.3 | 171.5 | 172.0 | 173.0 | 175.0 |
| IQ3_M | 168.61 GiB | 169.7 | 169.9 | 170.1 | 170.6 | 171.6 | 173.6 |
| Q3_K_S | 160.91 GiB | 162.0 | 162.2 | 162.4 | 162.9 | 163.9 | 165.9 |
| IQ3_XS | 151.27 GiB | 152.3 | 152.5 | 152.8 | 153.3 | 154.3 | 156.3 |
| IQ3_XXS | 146.20 GiB | 147.3 | 147.4 | 147.7 | 148.2 | 149.2 | 151.2 |
| UD-IQ2_M | 143.68 GiB | 144.8 | 144.9 | 145.2 | 145.7 | 146.7 | 148.7 |
| UD-IQ2_XXS | 142.94 GiB | 144.0 | 144.2 | 144.4 | 144.9 | 145.9 | 147.9 |
| Q2_K_L | 135.17 GiB | 136.2 | 136.4 | 136.7 | 137.2 | 138.2 | 140.2 |
| Q2_K | 135.04 GiB | 136.1 | 136.3 | 136.5 | 137.0 | 138.0 | 140.0 |
| IQ2_M | 116.50 GiB | 117.6 | 117.7 | 118.0 | 118.5 | 119.5 | 121.5 |
| IQ2_S | 102.49 GiB | 103.6 | 103.7 | 104.0 | 104.5 | 105.5 | 107.5 |
| IQ2_XS | 102.27 GiB | 103.3 | 103.5 | 103.8 | 104.3 | 105.3 | 107.3 |
| IQ2_XXS | 88.58 GiB | 89.7 | 89.8 | 90.1 | 90.6 | 91.6 | 93.6 |
| IQ1_M | 79.29 GiB | 80.4 | 80.5 | 80.8 | 81.3 | 82.3 | 84.3 |
| IQ1_S | 76.06 GiB | 77.1 | 77.3 | 77.6 | 78.1 | 79.0 | 81.0 |
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.