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 q4_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 | 201.9 | 202.1 | 202.5 | 203.4 | 205.1 | 208.4 |
| Q8_0 | 106.67 GiB | 107.8 | 108.0 | 108.4 | 109.3 | 111.0 | 114.3 |
| Q6_K_L | 83.13 GiB | 84.3 | 84.5 | 84.9 | 85.7 | 87.4 | 90.8 |
| Q6_K | 82.67 GiB | 83.8 | 84.0 | 84.4 | 85.3 | 87.0 | 90.3 |
| Q5_K_L | 73.87 GiB | 75.0 | 75.2 | 75.6 | 76.5 | 78.2 | 81.5 |
| Q5_K_M | 71.29 GiB | 72.4 | 72.6 | 73.1 | 73.9 | 75.6 | 79.0 |
| Q5_K_S | 69.16 GiB | 70.3 | 70.5 | 70.9 | 71.8 | 73.5 | 76.8 |
| Q4_1 | 64.35 GiB | 65.5 | 65.7 | 66.1 | 67.0 | 68.7 | 72.0 |
| Q4_K_L | 63.62 GiB | 64.8 | 65.0 | 65.4 | 66.2 | 67.9 | 71.3 |
| Q4_K_M | 62.91 GiB | 64.0 | 64.3 | 64.7 | 65.5 | 67.2 | 70.6 |
| Q4_0 | 58.72 GiB | 59.9 | 60.1 | 60.5 | 61.3 | 63.0 | 66.4 |
| IQ4_NL | 58.67 GiB | 59.8 | 60.0 | 60.4 | 61.3 | 63.0 | 66.3 |
| Q4_K_S | 57.23 GiB | 58.4 | 58.6 | 59.0 | 59.8 | 61.5 | 64.9 |
| IQ4_XS | 55.78 GiB | 56.9 | 57.1 | 57.5 | 58.4 | 60.1 | 63.5 |
| Q3_K_L | 53.83 GiB | 55.0 | 55.2 | 55.6 | 56.4 | 58.1 | 61.5 |
| Q3_K_M | 50.59 GiB | 51.7 | 51.9 | 52.4 | 53.2 | 54.9 | 58.3 |
| IQ3_M | 46.87 GiB | 48.0 | 48.2 | 48.6 | 49.5 | 51.2 | 54.5 |
| Q3_K_S | 46.34 GiB | 47.5 | 47.7 | 48.1 | 49.0 | 50.6 | 54.0 |
| IQ3_XS | 44.19 GiB | 45.3 | 45.5 | 46.0 | 46.8 | 48.5 | 51.9 |
| UD-IQ3_XXS | 42.59 GiB | 43.7 | 43.9 | 44.4 | 45.2 | 46.9 | 50.3 |
| IQ3_XXS | 41.87 GiB | 43.0 | 43.2 | 43.6 | 44.5 | 46.2 | 49.5 |
| Q2_K_L | 40.97 GiB | 42.1 | 42.3 | 42.7 | 43.6 | 45.3 | 48.7 |
| Q2_K | 40.03 GiB | 41.2 | 41.4 | 41.8 | 42.6 | 44.3 | 47.7 |
| UD-IQ2_M | 36.39 GiB | 37.5 | 37.7 | 38.2 | 39.0 | 40.7 | 44.1 |
| UD-IQ2_XXS | 34.83 GiB | 36.0 | 36.2 | 36.6 | 37.4 | 39.1 | 42.5 |
| IQ2_M | 34.56 GiB | 35.7 | 35.9 | 36.3 | 37.2 | 38.9 | 42.2 |
| UD-IQ1_M | 32.59 GiB | 33.7 | 33.9 | 34.4 | 35.2 | 36.9 | 40.3 |
| IQ2_S | 31.98 GiB | 33.1 | 33.3 | 33.7 | 34.6 | 36.3 | 39.7 |
| IQ2_XS | 30.68 GiB | 31.8 | 32.0 | 32.4 | 33.3 | 35.0 | 38.4 |
| UD-IQ1_S | 30.24 GiB | 31.4 | 31.6 | 32.0 | 32.9 | 34.5 | 37.9 |
| IQ2_XXS | 28.09 GiB | 29.2 | 29.4 | 29.9 | 30.7 | 32.4 | 35.8 |
| UD-TQ1_0 | 27.25 GiB | 28.4 | 28.6 | 29.0 | 29.9 | 31.6 | 34.9 |
| IQ1_M | 24.51 GiB | 25.6 | 25.9 | 26.3 | 27.1 | 28.8 | 32.2 |
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.