Can I run Llama-4-Scout-17B-16E-Instruct on a Radeon RX 6500 XT?
Not at these settings. No indexed quantization of Llama-4-Scout-17B-16E-Instruct fits Radeon RX 6500 XT at any context we compute, with q8_0 KV. The smallest shipped quantization is 24.51 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 | 200.76 GiB | 202.1 | 202.5 | 203.3 | 204.9 | 208.1 | 214.4 |
| Q8_0 | 106.67 GiB | 108.0 | 108.4 | 109.2 | 110.8 | 114.0 | 120.3 |
| Q6_K_L | 83.13 GiB | 84.5 | 84.9 | 85.7 | 87.2 | 90.4 | 96.8 |
| Q6_K | 82.67 GiB | 84.0 | 84.4 | 85.2 | 86.8 | 90.0 | 96.3 |
| Q5_K_L | 73.87 GiB | 75.2 | 75.6 | 76.4 | 78.0 | 81.2 | 87.5 |
| Q5_K_M | 71.29 GiB | 72.6 | 73.0 | 73.8 | 75.4 | 78.6 | 85.0 |
| Q5_K_S | 69.16 GiB | 70.5 | 70.9 | 71.7 | 73.3 | 76.5 | 82.8 |
| Q4_1 | 64.35 GiB | 65.7 | 66.1 | 66.9 | 68.5 | 71.7 | 78.0 |
| Q4_K_L | 63.62 GiB | 64.9 | 65.3 | 66.1 | 67.7 | 70.9 | 77.3 |
| Q4_K_M | 62.91 GiB | 64.2 | 64.6 | 65.4 | 67.0 | 70.2 | 76.6 |
| Q4_0 | 58.72 GiB | 60.0 | 60.4 | 61.2 | 62.8 | 66.0 | 72.4 |
| IQ4_NL | 58.67 GiB | 60.0 | 60.4 | 61.2 | 62.8 | 66.0 | 72.3 |
| Q4_K_S | 57.23 GiB | 58.6 | 59.0 | 59.8 | 61.3 | 64.5 | 70.9 |
| IQ4_XS | 55.78 GiB | 57.1 | 57.5 | 58.3 | 59.9 | 63.1 | 69.5 |
| Q3_K_L | 53.83 GiB | 55.2 | 55.6 | 56.4 | 57.9 | 61.1 | 67.5 |
| Q3_K_M | 50.59 GiB | 51.9 | 52.3 | 53.1 | 54.7 | 57.9 | 64.3 |
| IQ3_M | 46.87 GiB | 48.2 | 48.6 | 49.4 | 51.0 | 54.2 | 60.5 |
| Q3_K_S | 46.34 GiB | 47.7 | 48.1 | 48.9 | 50.5 | 53.6 | 60.0 |
| IQ3_XS | 44.19 GiB | 45.5 | 45.9 | 46.7 | 48.3 | 51.5 | 57.9 |
| UD-IQ3_XXS | 42.59 GiB | 43.9 | 44.3 | 45.1 | 46.7 | 49.9 | 56.3 |
| IQ3_XXS | 41.87 GiB | 43.2 | 43.6 | 44.4 | 46.0 | 49.2 | 55.5 |
| Q2_K_L | 40.97 GiB | 42.3 | 42.7 | 43.5 | 45.1 | 48.3 | 54.7 |
| Q2_K | 40.03 GiB | 41.4 | 41.8 | 42.6 | 44.1 | 47.3 | 53.7 |
| UD-IQ2_M | 36.39 GiB | 37.7 | 38.1 | 38.9 | 40.5 | 43.7 | 50.1 |
| UD-IQ2_XXS | 34.83 GiB | 36.2 | 36.6 | 37.4 | 38.9 | 42.1 | 48.5 |
| IQ2_M | 34.56 GiB | 35.9 | 36.3 | 37.1 | 38.7 | 41.9 | 48.2 |
| UD-IQ1_M | 32.59 GiB | 33.9 | 34.3 | 35.1 | 36.7 | 39.9 | 46.3 |
| IQ2_S | 31.98 GiB | 33.3 | 33.7 | 34.5 | 36.1 | 39.3 | 45.7 |
| IQ2_XS | 30.68 GiB | 32.0 | 32.4 | 33.2 | 34.8 | 38.0 | 44.4 |
| UD-IQ1_S | 30.24 GiB | 31.6 | 32.0 | 32.8 | 34.4 | 37.5 | 43.9 |
| IQ2_XXS | 28.09 GiB | 29.4 | 29.8 | 30.6 | 32.2 | 35.4 | 41.8 |
| UD-TQ1_0 | 27.25 GiB | 28.6 | 29.0 | 29.8 | 31.4 | 34.6 | 40.9 |
| IQ1_M | 24.51 GiB | 25.8 | 26.2 | 27.0 | 28.6 | 31.8 | 38.2 |
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 8,192-token window rather than the full context.