Can I run Trinity-Large-TrueBase on a GeForce RTX 5070?
Not at these settings. No indexed quantization of Trinity-Large-TrueBase fits GeForce RTX 5070 at any context we compute, with q8_0 KV. The smallest shipped quantization is 73.49 GiB in weights alone, against 11.16 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 | 395.9 | 396.1 | 396.3 | 396.8 | 397.8 | 399.8 |
| Q6_K | 305.07 GiB | 306.4 | 306.6 | 306.8 | 307.3 | 308.3 | 310.3 |
| Q5_K_M | 263.87 GiB | 265.2 | 265.4 | 265.6 | 266.1 | 267.1 | 269.1 |
| Q5_K_S | 255.88 GiB | 257.2 | 257.4 | 257.6 | 258.1 | 259.1 | 261.1 |
| Q4_1 | 232.76 GiB | 234.1 | 234.3 | 234.5 | 235.0 | 236.0 | 238.0 |
| Q4_K_L | 225.46 GiB | 226.8 | 227.0 | 227.2 | 227.7 | 228.7 | 230.7 |
| Q4_K_M | 225.04 GiB | 226.4 | 226.5 | 226.8 | 227.3 | 228.3 | 230.3 |
| Q4_K_S | 217.12 GiB | 218.4 | 218.6 | 218.9 | 219.4 | 220.4 | 222.4 |
| Q4_0 | 213.32 GiB | 214.6 | 214.8 | 215.1 | 215.6 | 216.6 | 218.5 |
| IQ4_NL | 209.71 GiB | 211.0 | 211.2 | 211.5 | 212.0 | 212.9 | 214.9 |
| IQ4_XS | 198.19 GiB | 199.5 | 199.7 | 199.9 | 200.4 | 201.4 | 203.4 |
| I1-Q3_K_M | 176.30 GiB | 177.6 | 177.8 | 178.0 | 178.5 | 179.5 | 181.5 |
| Q3_K_L | 175.82 GiB | 177.1 | 177.3 | 177.6 | 178.1 | 179.1 | 181.0 |
| Q3_K_M | 168.74 GiB | 170.1 | 170.2 | 170.5 | 171.0 | 172.0 | 174.0 |
| IQ3_M | 168.61 GiB | 169.9 | 170.1 | 170.4 | 170.9 | 171.9 | 173.8 |
| Q3_K_S | 160.91 GiB | 162.2 | 162.4 | 162.7 | 163.1 | 164.1 | 166.1 |
| I1-IQ3_M | 160.39 GiB | 161.7 | 161.9 | 162.1 | 162.6 | 163.6 | 165.6 |
| I1-IQ3_S | 159.92 GiB | 161.2 | 161.4 | 161.7 | 162.2 | 163.2 | 165.2 |
| I1-Q3_K_S | 159.90 GiB | 161.2 | 161.4 | 161.6 | 162.1 | 163.1 | 165.1 |
| IQ3_XS | 151.27 GiB | 152.6 | 152.8 | 153.0 | 153.5 | 154.5 | 156.5 |
| I1-IQ3_XS | 150.34 GiB | 151.7 | 151.8 | 152.1 | 152.6 | 153.6 | 155.6 |
| IQ3_XXS | 146.20 GiB | 147.5 | 147.7 | 147.9 | 148.4 | 149.4 | 151.4 |
| I1-IQ3_XXS | 142.48 GiB | 143.8 | 144.0 | 144.2 | 144.7 | 145.7 | 147.7 |
| I1-Q2_K | 134.81 GiB | 136.1 | 136.3 | 136.6 | 137.1 | 138.1 | 140.0 |
| Q2_K_L | 130.00 GiB | 131.3 | 131.5 | 131.7 | 132.2 | 133.2 | 135.2 |
| Q2_K | 129.44 GiB | 130.8 | 130.9 | 131.2 | 131.7 | 132.7 | 134.7 |
| I1-Q2_K_S | 122.88 GiB | 124.2 | 124.4 | 124.6 | 125.1 | 126.1 | 128.1 |
| I1-IQ2_M | 119.77 GiB | 121.1 | 121.3 | 121.5 | 122.0 | 123.0 | 125.0 |
| IQ2_M | 116.50 GiB | 117.8 | 118.0 | 118.2 | 118.7 | 119.7 | 121.7 |
| I1-IQ2_S | 108.32 GiB | 109.6 | 109.8 | 110.1 | 110.6 | 111.6 | 113.6 |
| I1-IQ2_XS | 107.87 GiB | 109.2 | 109.4 | 109.6 | 110.1 | 111.1 | 113.1 |
| IQ2_S | 102.49 GiB | 103.8 | 104.0 | 104.2 | 104.7 | 105.7 | 107.7 |
| IQ2_XS | 102.27 GiB | 103.6 | 103.8 | 104.0 | 104.5 | 105.5 | 107.5 |
| I1-IQ2_XXS | 96.39 GiB | 97.7 | 97.9 | 98.1 | 98.6 | 99.6 | 101.6 |
| IQ2_XXS | 88.58 GiB | 89.9 | 90.1 | 90.3 | 90.8 | 91.8 | 93.8 |
| I1-IQ1_M | 82.07 GiB | 83.4 | 83.6 | 83.8 | 84.3 | 85.3 | 87.3 |
| IQ1_M | 79.29 GiB | 80.6 | 80.8 | 81.0 | 81.5 | 82.5 | 84.5 |
| IQ1_S | 76.06 GiB | 77.4 | 77.6 | 77.8 | 78.3 | 79.3 | 81.3 |
| I1-IQ1_S | 73.49 GiB | 74.8 | 75.0 | 75.2 | 75.7 | 76.7 | 78.7 |
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.