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 q4_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 | 395.4 | 395.5 | 395.6 | 395.9 | 396.4 | 397.5 |
| Q6_K_L | 319.94 GiB | 320.8 | 320.9 | 321.0 | 321.3 | 321.8 | 322.8 |
| Q6_K | 319.66 GiB | 320.5 | 320.6 | 320.7 | 321.0 | 321.5 | 322.6 |
| Q5_K_L | 264.48 GiB | 265.3 | 265.4 | 265.5 | 265.8 | 266.3 | 267.4 |
| Q5_K_M | 264.13 GiB | 265.0 | 265.1 | 265.2 | 265.5 | 266.0 | 267.0 |
| Q5_K_S | 256.01 GiB | 256.9 | 256.9 | 257.1 | 257.3 | 257.9 | 258.9 |
| Q4_1 | 232.94 GiB | 233.8 | 233.9 | 234.0 | 234.3 | 234.8 | 235.8 |
| Q4_K_L | 225.66 GiB | 226.5 | 226.6 | 226.7 | 227.0 | 227.5 | 228.6 |
| Q4_K_M | 225.24 GiB | 226.1 | 226.2 | 226.3 | 226.6 | 227.1 | 228.1 |
| Q4_K_S | 217.21 GiB | 218.1 | 218.1 | 218.3 | 218.5 | 219.1 | 220.1 |
| Q4_0 | 210.73 GiB | 211.6 | 211.7 | 211.8 | 212.1 | 212.6 | 213.6 |
| IQ4_NL | 209.91 GiB | 210.8 | 210.8 | 211.0 | 211.2 | 211.8 | 212.8 |
| IQ4_XS | 198.41 GiB | 199.3 | 199.3 | 199.5 | 199.7 | 200.3 | 201.3 |
| IQ3_M | 176.24 GiB | 177.1 | 177.2 | 177.3 | 177.6 | 178.1 | 179.1 |
| Q3_K_L | 176.09 GiB | 176.9 | 177.0 | 177.2 | 177.4 | 177.9 | 179.0 |
| Q3_K_M | 168.90 GiB | 169.7 | 169.8 | 170.0 | 170.2 | 170.8 | 171.8 |
| IQ3_XS | 168.77 GiB | 169.6 | 169.7 | 169.8 | 170.1 | 170.6 | 171.7 |
| Q3_K_S | 161.09 GiB | 161.9 | 162.0 | 162.2 | 162.4 | 162.9 | 164.0 |
| IQ3_XXS | 154.25 GiB | 155.1 | 155.2 | 155.3 | 155.6 | 156.1 | 157.2 |
| Q2_K_L | 130.22 GiB | 131.1 | 131.1 | 131.3 | 131.5 | 132.1 | 133.1 |
| Q2_K | 129.66 GiB | 130.5 | 130.6 | 130.7 | 131.0 | 131.5 | 132.6 |
| IQ2_M | 123.88 GiB | 124.7 | 124.8 | 124.9 | 125.2 | 125.7 | 126.8 |
| IQ2_S | 112.03 GiB | 112.9 | 113.0 | 113.1 | 113.4 | 113.9 | 114.9 |
| IQ2_XS | 110.12 GiB | 111.0 | 111.1 | 111.2 | 111.4 | 112.0 | 113.0 |
| IQ2_XXS | 98.70 GiB | 99.5 | 99.6 | 99.8 | 100.0 | 100.6 | 101.6 |
| IQ1_M | 84.63 GiB | 85.5 | 85.6 | 85.7 | 86.0 | 86.5 | 87.5 |
| IQ1_S | 75.64 GiB | 76.5 | 76.6 | 76.7 | 77.0 | 77.5 | 78.5 |
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 4,096-token window rather than the full context.