Can I run Llama-4-Scout-17B-16E-Instruct on a GeForce RTX 5090 D V2?
Not at these settings. No indexed quantization of Llama-4-Scout-17B-16E-Instruct fits GeForce RTX 5090 D V2 at any context we compute, with q8_0 KV. The smallest shipped quantization is 24.51 GiB in weights alone, against 22.32 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 | 202.0 | 202.4 | 203.2 | 204.8 | 208.0 | 214.3 |
| Q8_0 | 106.67 GiB | 107.9 | 108.3 | 109.1 | 110.7 | 113.9 | 120.2 |
| Q6_K_L | 83.13 GiB | 84.4 | 84.8 | 85.6 | 87.1 | 90.3 | 96.7 |
| Q6_K | 82.67 GiB | 83.9 | 84.3 | 85.1 | 86.7 | 89.9 | 96.2 |
| Q5_K_L | 73.87 GiB | 75.1 | 75.5 | 76.3 | 77.9 | 81.1 | 87.4 |
| Q5_K_M | 71.29 GiB | 72.5 | 72.9 | 73.7 | 75.3 | 78.5 | 84.9 |
| Q5_K_S | 69.16 GiB | 70.4 | 70.8 | 71.6 | 73.2 | 76.4 | 82.7 |
| Q4_1 | 64.35 GiB | 65.6 | 66.0 | 66.8 | 68.4 | 71.6 | 77.9 |
| Q4_K_L | 63.62 GiB | 64.8 | 65.2 | 66.0 | 67.6 | 70.8 | 77.2 |
| Q4_K_M | 62.91 GiB | 64.1 | 64.5 | 65.3 | 66.9 | 70.1 | 76.5 |
| Q4_0 | 58.72 GiB | 59.9 | 60.3 | 61.1 | 62.7 | 65.9 | 72.3 |
| IQ4_NL | 58.67 GiB | 59.9 | 60.3 | 61.1 | 62.7 | 65.9 | 72.2 |
| Q4_K_S | 57.23 GiB | 58.5 | 58.9 | 59.7 | 61.2 | 64.4 | 70.8 |
| IQ4_XS | 55.78 GiB | 57.0 | 57.4 | 58.2 | 59.8 | 63.0 | 69.4 |
| Q3_K_L | 53.83 GiB | 55.1 | 55.5 | 56.3 | 57.8 | 61.0 | 67.4 |
| Q3_K_M | 50.59 GiB | 51.8 | 52.2 | 53.0 | 54.6 | 57.8 | 64.2 |
| IQ3_M | 46.87 GiB | 48.1 | 48.5 | 49.3 | 50.9 | 54.1 | 60.4 |
| Q3_K_S | 46.34 GiB | 47.6 | 48.0 | 48.8 | 50.4 | 53.5 | 59.9 |
| IQ3_XS | 44.19 GiB | 45.4 | 45.8 | 46.6 | 48.2 | 51.4 | 57.8 |
| UD-IQ3_XXS | 42.59 GiB | 43.8 | 44.2 | 45.0 | 46.6 | 49.8 | 56.2 |
| IQ3_XXS | 41.87 GiB | 43.1 | 43.5 | 44.3 | 45.9 | 49.1 | 55.4 |
| Q2_K_L | 40.97 GiB | 42.2 | 42.6 | 43.4 | 45.0 | 48.2 | 54.6 |
| Q2_K | 40.03 GiB | 41.3 | 41.7 | 42.5 | 44.0 | 47.2 | 53.6 |
| UD-IQ2_M | 36.39 GiB | 37.6 | 38.0 | 38.8 | 40.4 | 43.6 | 50.0 |
| UD-IQ2_XXS | 34.83 GiB | 36.1 | 36.5 | 37.3 | 38.8 | 42.0 | 48.4 |
| IQ2_M | 34.56 GiB | 35.8 | 36.2 | 37.0 | 38.6 | 41.8 | 48.1 |
| UD-IQ1_M | 32.59 GiB | 33.8 | 34.2 | 35.0 | 36.6 | 39.8 | 46.2 |
| IQ2_S | 31.98 GiB | 33.2 | 33.6 | 34.4 | 36.0 | 39.2 | 45.6 |
| IQ2_XS | 30.68 GiB | 31.9 | 32.3 | 33.1 | 34.7 | 37.9 | 44.3 |
| UD-IQ1_S | 30.24 GiB | 31.5 | 31.9 | 32.7 | 34.3 | 37.4 | 43.8 |
| IQ2_XXS | 28.09 GiB | 29.3 | 29.7 | 30.5 | 32.1 | 35.3 | 41.7 |
| UD-TQ1_0 | 27.25 GiB | 28.5 | 28.9 | 29.7 | 31.3 | 34.5 | 40.8 |
| IQ1_M | 24.51 GiB | 25.7 | 26.1 | 26.9 | 28.5 | 31.7 | 38.1 |
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.