Can I run Trinity-Large-Thinking on a Apple M3 Ultra?
Not at these settings. No indexed quantization of Trinity-Large-Thinking fits Apple M3 Ultra at any context we compute, with q8_0 KV. The smallest shipped quantization is 75.64 GiB in weights alone, against 66.96 GiB usable. CPU offload can still run it, slowly.
Every quantization at every context
| Quant | Weights● | 4K◐ | 8K◐ | 16K◐ | 32K◐ | 64K◐ | 128K◐ |
|---|---|---|---|---|---|---|---|
| Q8_0 | 394.59 GiB | 395.7 | 395.8 | 396.1 | 396.6 | 397.6 | 399.6 |
| Q6_K_L | 319.94 GiB | 321.0 | 321.2 | 321.4 | 321.9 | 322.9 | 324.9 |
| Q6_K | 319.66 GiB | 320.7 | 320.9 | 321.2 | 321.7 | 322.7 | 324.6 |
| Q5_K_L | 264.48 GiB | 265.6 | 265.7 | 266.0 | 266.5 | 267.5 | 269.5 |
| Q5_K_M | 264.13 GiB | 265.2 | 265.4 | 265.6 | 266.1 | 267.1 | 269.1 |
| Q5_K_S | 256.01 GiB | 257.1 | 257.3 | 257.5 | 258.0 | 259.0 | 261.0 |
| Q4_1 | 232.94 GiB | 234.0 | 234.2 | 234.4 | 234.9 | 235.9 | 237.9 |
| Q4_K_L | 225.66 GiB | 226.7 | 226.9 | 227.2 | 227.7 | 228.7 | 230.6 |
| Q4_K_M | 225.24 GiB | 226.3 | 226.5 | 226.7 | 227.2 | 228.2 | 230.2 |
| Q4_K_S | 217.21 GiB | 218.3 | 218.5 | 218.7 | 219.2 | 220.2 | 222.2 |
| Q4_0 | 210.73 GiB | 211.8 | 212.0 | 212.2 | 212.7 | 213.7 | 215.7 |
| IQ4_NL | 209.91 GiB | 211.0 | 211.2 | 211.4 | 211.9 | 212.9 | 214.9 |
| IQ4_XS | 198.41 GiB | 199.5 | 199.7 | 199.9 | 200.4 | 201.4 | 203.4 |
| IQ3_M | 176.24 GiB | 177.3 | 177.5 | 177.7 | 178.2 | 179.2 | 181.2 |
| Q3_K_L | 176.09 GiB | 177.2 | 177.3 | 177.6 | 178.1 | 179.1 | 181.1 |
| Q3_K_M | 168.90 GiB | 170.0 | 170.1 | 170.4 | 170.9 | 171.9 | 173.9 |
| IQ3_XS | 168.77 GiB | 169.8 | 170.0 | 170.3 | 170.8 | 171.8 | 173.8 |
| Q3_K_S | 161.09 GiB | 162.2 | 162.3 | 162.6 | 163.1 | 164.1 | 166.1 |
| IQ3_XXS | 154.25 GiB | 155.3 | 155.5 | 155.7 | 156.2 | 157.2 | 159.2 |
| Q2_K_L | 130.22 GiB | 131.3 | 131.5 | 131.7 | 132.2 | 133.2 | 135.2 |
| Q2_K | 129.66 GiB | 130.7 | 130.9 | 131.2 | 131.7 | 132.6 | 134.6 |
| IQ2_M | 123.88 GiB | 125.0 | 125.1 | 125.4 | 125.9 | 126.9 | 128.9 |
| IQ2_S | 112.03 GiB | 113.1 | 113.3 | 113.5 | 114.0 | 115.0 | 117.0 |
| IQ2_XS | 110.12 GiB | 111.2 | 111.4 | 111.6 | 112.1 | 113.1 | 115.1 |
| IQ2_XXS | 98.70 GiB | 99.8 | 99.9 | 100.2 | 100.7 | 101.7 | 103.7 |
| IQ1_M | 84.63 GiB | 85.7 | 85.9 | 86.1 | 86.6 | 87.6 | 89.6 |
| IQ1_S | 75.64 GiB | 76.7 | 76.9 | 77.1 | 77.6 | 78.6 | 80.6 |
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.