Can I run Llama-4-Scout-17B-16E-Instruct on a GeForce RTX 3080 Ti?
Not at these settings. No indexed quantization of Llama-4-Scout-17B-16E-Instruct fits GeForce RTX 3080 Ti at any context we compute, with f16 KV. The smallest shipped quantization is 24.51 GiB in weights alone, against 18.60 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 | 202.3 | 203.1 | 204.6 | 207.6 | 213.6 | 225.6 |
| Q8_0 | 106.67 GiB | 108.2 | 109.0 | 110.5 | 113.5 | 119.5 | 131.5 |
| Q6_K_L | 83.13 GiB | 84.7 | 85.5 | 87.0 | 90.0 | 96.0 | 108.0 |
| Q6_K | 82.67 GiB | 84.2 | 85.0 | 86.5 | 89.5 | 95.5 | 107.5 |
| Q5_K_L | 73.87 GiB | 75.4 | 76.2 | 77.7 | 80.7 | 86.7 | 98.7 |
| Q5_K_M | 71.29 GiB | 72.9 | 73.6 | 75.1 | 78.1 | 84.1 | 96.1 |
| Q5_K_S | 69.16 GiB | 70.7 | 71.5 | 73.0 | 76.0 | 82.0 | 94.0 |
| Q4_1 | 64.35 GiB | 65.9 | 66.7 | 68.2 | 71.2 | 77.2 | 89.2 |
| Q4_K_L | 63.62 GiB | 65.2 | 65.9 | 67.4 | 70.4 | 76.4 | 88.4 |
| Q4_K_M | 62.91 GiB | 64.5 | 65.2 | 66.7 | 69.7 | 75.7 | 87.7 |
| Q4_0 | 58.72 GiB | 60.3 | 61.1 | 62.6 | 65.6 | 71.6 | 83.6 |
| IQ4_NL | 58.67 GiB | 60.2 | 61.0 | 62.5 | 65.5 | 71.5 | 83.5 |
| Q4_K_S | 57.23 GiB | 58.8 | 59.6 | 61.1 | 64.1 | 70.1 | 82.1 |
| IQ4_XS | 55.78 GiB | 57.4 | 58.1 | 59.6 | 62.6 | 68.6 | 80.6 |
| Q3_K_L | 53.83 GiB | 55.4 | 56.2 | 57.7 | 60.7 | 66.7 | 78.7 |
| Q3_K_M | 50.59 GiB | 52.2 | 52.9 | 54.4 | 57.4 | 63.4 | 75.4 |
| IQ3_M | 46.87 GiB | 48.4 | 49.2 | 50.7 | 53.7 | 59.7 | 71.7 |
| Q3_K_S | 46.34 GiB | 47.9 | 48.7 | 50.2 | 53.2 | 59.2 | 71.2 |
| IQ3_XS | 44.19 GiB | 45.8 | 46.5 | 48.0 | 51.0 | 57.0 | 69.0 |
| UD-IQ3_XXS | 42.59 GiB | 44.2 | 44.9 | 46.4 | 49.4 | 55.4 | 67.4 |
| IQ3_XXS | 41.87 GiB | 43.4 | 44.2 | 45.7 | 48.7 | 54.7 | 66.7 |
| Q2_K_L | 40.97 GiB | 42.6 | 43.3 | 44.8 | 47.8 | 53.8 | 65.8 |
| Q2_K | 40.03 GiB | 41.6 | 42.4 | 43.9 | 46.9 | 52.9 | 64.9 |
| UD-IQ2_M | 36.39 GiB | 38.0 | 38.7 | 40.2 | 43.2 | 49.2 | 61.2 |
| UD-IQ2_XXS | 34.83 GiB | 36.4 | 37.2 | 38.7 | 41.7 | 47.7 | 59.7 |
| IQ2_M | 34.56 GiB | 36.1 | 36.9 | 38.4 | 41.4 | 47.4 | 59.4 |
| UD-IQ1_M | 32.59 GiB | 34.2 | 34.9 | 36.4 | 39.4 | 45.4 | 57.4 |
| IQ2_S | 31.98 GiB | 33.6 | 34.3 | 35.8 | 38.8 | 44.8 | 56.8 |
| IQ2_XS | 30.68 GiB | 32.3 | 33.0 | 34.5 | 37.5 | 43.5 | 55.5 |
| UD-IQ1_S | 30.24 GiB | 31.8 | 32.6 | 34.1 | 37.1 | 43.1 | 55.1 |
| IQ2_XXS | 28.09 GiB | 29.7 | 30.4 | 31.9 | 34.9 | 40.9 | 52.9 |
| UD-TQ1_0 | 27.25 GiB | 28.8 | 29.6 | 31.1 | 34.1 | 40.1 | 52.1 |
| IQ1_M | 24.51 GiB | 26.1 | 26.8 | 28.3 | 31.3 | 37.3 | 49.3 |
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 8,192-token window rather than the full context.