Does Llama-4-Scout-17B-16E-Instruct fit in 16GB of VRAM?
Not at these settings. No indexed quantization of Llama-4-Scout-17B-16E-Instruct fits 16GB card at any context we compute, with q4_0 KV. The smallest shipped quantization is 24.51 GiB in weights alone, against 14.88 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 | 201.8 | 202.0 | 202.4 | 203.3 | 205.0 | 208.3 |
| Q8_0 | 106.67 GiB | 107.7 | 107.9 | 108.3 | 109.2 | 110.9 | 114.2 |
| Q6_K_L | 83.13 GiB | 84.2 | 84.4 | 84.8 | 85.6 | 87.3 | 90.7 |
| Q6_K | 82.67 GiB | 83.7 | 83.9 | 84.3 | 85.2 | 86.9 | 90.2 |
| Q5_K_L | 73.87 GiB | 74.9 | 75.1 | 75.5 | 76.4 | 78.1 | 81.4 |
| Q5_K_M | 71.29 GiB | 72.3 | 72.5 | 73.0 | 73.8 | 75.5 | 78.9 |
| Q5_K_S | 69.16 GiB | 70.2 | 70.4 | 70.8 | 71.7 | 73.4 | 76.7 |
| Q4_1 | 64.35 GiB | 65.4 | 65.6 | 66.0 | 66.9 | 68.6 | 71.9 |
| Q4_K_L | 63.62 GiB | 64.7 | 64.9 | 65.3 | 66.1 | 67.8 | 71.2 |
| Q4_K_M | 62.91 GiB | 63.9 | 64.2 | 64.6 | 65.4 | 67.1 | 70.5 |
| Q4_0 | 58.72 GiB | 59.8 | 60.0 | 60.4 | 61.2 | 62.9 | 66.3 |
| IQ4_NL | 58.67 GiB | 59.7 | 59.9 | 60.3 | 61.2 | 62.9 | 66.2 |
| Q4_K_S | 57.23 GiB | 58.3 | 58.5 | 58.9 | 59.7 | 61.4 | 64.8 |
| IQ4_XS | 55.78 GiB | 56.8 | 57.0 | 57.4 | 58.3 | 60.0 | 63.4 |
| Q3_K_L | 53.83 GiB | 54.9 | 55.1 | 55.5 | 56.3 | 58.0 | 61.4 |
| Q3_K_M | 50.59 GiB | 51.6 | 51.8 | 52.3 | 53.1 | 54.8 | 58.2 |
| IQ3_M | 46.87 GiB | 47.9 | 48.1 | 48.5 | 49.4 | 51.1 | 54.4 |
| Q3_K_S | 46.34 GiB | 47.4 | 47.6 | 48.0 | 48.9 | 50.5 | 53.9 |
| IQ3_XS | 44.19 GiB | 45.2 | 45.4 | 45.9 | 46.7 | 48.4 | 51.8 |
| UD-IQ3_XXS | 42.59 GiB | 43.6 | 43.8 | 44.3 | 45.1 | 46.8 | 50.2 |
| IQ3_XXS | 41.87 GiB | 42.9 | 43.1 | 43.5 | 44.4 | 46.1 | 49.4 |
| Q2_K_L | 40.97 GiB | 42.0 | 42.2 | 42.6 | 43.5 | 45.2 | 48.6 |
| Q2_K | 40.03 GiB | 41.1 | 41.3 | 41.7 | 42.5 | 44.2 | 47.6 |
| UD-IQ2_M | 36.39 GiB | 37.4 | 37.6 | 38.1 | 38.9 | 40.6 | 44.0 |
| UD-IQ2_XXS | 34.83 GiB | 35.9 | 36.1 | 36.5 | 37.3 | 39.0 | 42.4 |
| IQ2_M | 34.56 GiB | 35.6 | 35.8 | 36.2 | 37.1 | 38.8 | 42.1 |
| UD-IQ1_M | 32.59 GiB | 33.6 | 33.8 | 34.3 | 35.1 | 36.8 | 40.2 |
| IQ2_S | 31.98 GiB | 33.0 | 33.2 | 33.6 | 34.5 | 36.2 | 39.6 |
| IQ2_XS | 30.68 GiB | 31.7 | 31.9 | 32.3 | 33.2 | 34.9 | 38.3 |
| UD-IQ1_S | 30.24 GiB | 31.3 | 31.5 | 31.9 | 32.8 | 34.4 | 37.8 |
| IQ2_XXS | 28.09 GiB | 29.1 | 29.3 | 29.8 | 30.6 | 32.3 | 35.7 |
| UD-TQ1_0 | 27.25 GiB | 28.3 | 28.5 | 28.9 | 29.8 | 31.5 | 34.8 |
| IQ1_M | 24.51 GiB | 25.5 | 25.8 | 26.2 | 27.0 | 28.7 | 32.1 |
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 there are no speeds on this page
A capacity is not a card. Whether a model fits depends only on memory, so every figure above holds for any 16GB accelerator. How fast it runs depends on memory bandwidth, which varies several-fold between cards of the same capacity — so putting a tokens-per-second number here would be inventing one. Pick a specific card from hardware and the speed column appears.
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.