Can I run Trinity-Large-Preview on a Radeon RX 6500 XT?
Not at these settings. No indexed quantization of Trinity-Large-Preview fits Radeon RX 6500 XT at any context we compute, with q4_0 KV. The smallest shipped quantization is 76.06 GiB in weights alone, against 3.72 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 | 743.8 | 743.9 | 744.0 | 744.3 | 744.8 | 745.9 |
| Q8_0 | 394.59 GiB | 395.8 | 395.9 | 396.0 | 396.3 | 396.8 | 397.8 |
| Q6_K_L | 305.35 GiB | 306.5 | 306.6 | 306.8 | 307.0 | 307.6 | 308.6 |
| Q6_K | 305.07 GiB | 306.3 | 306.4 | 306.5 | 306.7 | 307.3 | 308.3 |
| Q5_K_L | 264.22 GiB | 265.4 | 265.5 | 265.6 | 265.9 | 266.4 | 267.5 |
| Q5_K_M | 263.87 GiB | 265.1 | 265.1 | 265.3 | 265.5 | 266.1 | 267.1 |
| Q5_K_S | 255.88 GiB | 257.1 | 257.2 | 257.3 | 257.6 | 258.1 | 259.1 |
| Q4_1 | 232.76 GiB | 234.0 | 234.0 | 234.2 | 234.4 | 235.0 | 236.0 |
| Q4_K_L | 225.46 GiB | 226.7 | 226.7 | 226.9 | 227.1 | 227.7 | 228.7 |
| Q4_K_M | 225.04 GiB | 226.2 | 226.3 | 226.4 | 226.7 | 227.2 | 228.3 |
| Q4_K_S | 217.12 GiB | 218.3 | 218.4 | 218.5 | 218.8 | 219.3 | 220.4 |
| Q4_0 | 213.32 GiB | 214.5 | 214.6 | 214.7 | 215.0 | 215.5 | 216.6 |
| IQ4_NL | 209.71 GiB | 210.9 | 211.0 | 211.1 | 211.4 | 211.9 | 213.0 |
| IQ4_XS | 198.19 GiB | 199.4 | 199.5 | 199.6 | 199.9 | 200.4 | 201.5 |
| Q3_K_M | 176.43 GiB | 177.6 | 177.7 | 177.8 | 178.1 | 178.6 | 179.7 |
| Q3_K_L | 175.82 GiB | 177.0 | 177.1 | 177.2 | 177.5 | 178.0 | 179.1 |
| UD-IQ3_XXS | 170.01 GiB | 171.2 | 171.3 | 171.4 | 171.7 | 172.2 | 173.3 |
| IQ3_M | 168.61 GiB | 169.8 | 169.9 | 170.0 | 170.3 | 170.8 | 171.9 |
| Q3_K_S | 160.91 GiB | 162.1 | 162.2 | 162.3 | 162.6 | 163.1 | 164.2 |
| IQ3_XS | 151.27 GiB | 152.5 | 152.6 | 152.7 | 152.9 | 153.5 | 154.5 |
| IQ3_XXS | 146.20 GiB | 147.4 | 147.5 | 147.6 | 147.9 | 148.4 | 149.5 |
| UD-IQ2_M | 143.68 GiB | 144.9 | 145.0 | 145.1 | 145.4 | 145.9 | 146.9 |
| UD-IQ2_XXS | 142.94 GiB | 144.1 | 144.2 | 144.4 | 144.6 | 145.1 | 146.2 |
| Q2_K_L | 135.17 GiB | 136.4 | 136.5 | 136.6 | 136.8 | 137.4 | 138.4 |
| Q2_K | 135.04 GiB | 136.2 | 136.3 | 136.4 | 136.7 | 137.2 | 138.3 |
| IQ2_M | 116.50 GiB | 117.7 | 117.8 | 117.9 | 118.2 | 118.7 | 119.8 |
| IQ2_S | 102.49 GiB | 103.7 | 103.8 | 103.9 | 104.2 | 104.7 | 105.8 |
| IQ2_XS | 102.27 GiB | 103.5 | 103.6 | 103.7 | 103.9 | 104.5 | 105.5 |
| IQ2_XXS | 88.58 GiB | 89.8 | 89.9 | 90.0 | 90.3 | 90.8 | 91.8 |
| IQ1_M | 79.29 GiB | 80.5 | 80.6 | 80.7 | 81.0 | 81.5 | 82.5 |
| IQ1_S | 76.06 GiB | 77.2 | 77.3 | 77.5 | 77.7 | 78.3 | 79.3 |
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.