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 q8_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 | 744.0 | 744.2 | 744.4 | 744.9 | 745.9 | 747.9 |
| Q8_0 | 394.59 GiB | 396.0 | 396.2 | 396.4 | 396.9 | 397.9 | 399.9 |
| Q6_K_L | 305.35 GiB | 306.8 | 306.9 | 307.2 | 307.7 | 308.7 | 310.7 |
| Q6_K | 305.07 GiB | 306.5 | 306.7 | 306.9 | 307.4 | 308.4 | 310.4 |
| Q5_K_L | 264.22 GiB | 265.6 | 265.8 | 266.1 | 266.6 | 267.6 | 269.5 |
| Q5_K_M | 263.87 GiB | 265.3 | 265.5 | 265.7 | 266.2 | 267.2 | 269.2 |
| Q5_K_S | 255.88 GiB | 257.3 | 257.5 | 257.7 | 258.2 | 259.2 | 261.2 |
| Q4_1 | 232.76 GiB | 234.2 | 234.4 | 234.6 | 235.1 | 236.1 | 238.1 |
| Q4_K_L | 225.46 GiB | 226.9 | 227.1 | 227.3 | 227.8 | 228.8 | 230.8 |
| Q4_K_M | 225.04 GiB | 226.5 | 226.6 | 226.9 | 227.4 | 228.4 | 230.4 |
| Q4_K_S | 217.12 GiB | 218.5 | 218.7 | 219.0 | 219.5 | 220.5 | 222.5 |
| Q4_0 | 213.32 GiB | 214.7 | 214.9 | 215.2 | 215.7 | 216.7 | 218.6 |
| IQ4_NL | 209.71 GiB | 211.1 | 211.3 | 211.6 | 212.1 | 213.0 | 215.0 |
| IQ4_XS | 198.19 GiB | 199.6 | 199.8 | 200.0 | 200.5 | 201.5 | 203.5 |
| Q3_K_M | 176.43 GiB | 177.9 | 178.0 | 178.3 | 178.8 | 179.8 | 181.8 |
| Q3_K_L | 175.82 GiB | 177.2 | 177.4 | 177.7 | 178.2 | 179.2 | 181.1 |
| UD-IQ3_XXS | 170.01 GiB | 171.4 | 171.6 | 171.9 | 172.4 | 173.3 | 175.3 |
| IQ3_M | 168.61 GiB | 170.0 | 170.2 | 170.5 | 171.0 | 172.0 | 173.9 |
| Q3_K_S | 160.91 GiB | 162.3 | 162.5 | 162.8 | 163.2 | 164.2 | 166.2 |
| IQ3_XS | 151.27 GiB | 152.7 | 152.9 | 153.1 | 153.6 | 154.6 | 156.6 |
| IQ3_XXS | 146.20 GiB | 147.6 | 147.8 | 148.0 | 148.5 | 149.5 | 151.5 |
| UD-IQ2_M | 143.68 GiB | 145.1 | 145.3 | 145.5 | 146.0 | 147.0 | 149.0 |
| UD-IQ2_XXS | 142.94 GiB | 144.4 | 144.5 | 144.8 | 145.3 | 146.3 | 148.3 |
| Q2_K_L | 135.17 GiB | 136.6 | 136.8 | 137.0 | 137.5 | 138.5 | 140.5 |
| Q2_K | 135.04 GiB | 136.5 | 136.6 | 136.9 | 137.4 | 138.4 | 140.4 |
| IQ2_M | 116.50 GiB | 117.9 | 118.1 | 118.3 | 118.8 | 119.8 | 121.8 |
| IQ2_S | 102.49 GiB | 103.9 | 104.1 | 104.3 | 104.8 | 105.8 | 107.8 |
| IQ2_XS | 102.27 GiB | 103.7 | 103.9 | 104.1 | 104.6 | 105.6 | 107.6 |
| IQ2_XXS | 88.58 GiB | 90.0 | 90.2 | 90.4 | 90.9 | 91.9 | 93.9 |
| IQ1_M | 79.29 GiB | 80.7 | 80.9 | 81.1 | 81.6 | 82.6 | 84.6 |
| IQ1_S | 76.06 GiB | 77.5 | 77.7 | 77.9 | 78.4 | 79.4 | 81.4 |
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.