Can I run Laguna-S-2.1 on a GeForce RTX 3080?
Not at these settings. No indexed quantization of Laguna-S-2.1 fits GeForce RTX 3080 at any context we compute, with q8_0 KV. The smallest shipped quantization is 23.15 GiB in weights alone, against 9.30 GiB usable. CPU offload can still run it, slowly.
Every quantization at every context
| Quant | Weights● | 4K◐ | 8K◐ | 16K◐ | 32K◐ | 64K◐ | 128K◐ |
|---|---|---|---|---|---|---|---|
| BF16 | 219.05 GiB | 220.0 | 220.1 | 220.3 | 220.7 | 221.5 | 223.1 |
| F16 | 219.05 GiB | 220.0 | 220.1 | 220.3 | 220.7 | 221.5 | 223.1 |
| Q8_0 | 119.91 GiB | 120.9 | 121.0 | 121.2 | 121.6 | 122.4 | 124.0 |
| Q6_K_L | 94.83 GiB | 95.8 | 95.9 | 96.1 | 96.5 | 97.3 | 98.9 |
| UD-Q6_K | 91.19 GiB | 92.2 | 92.3 | 92.5 | 92.9 | 93.7 | 95.3 |
| Q6_K | 89.93 GiB | 90.9 | 91.0 | 91.2 | 91.6 | 92.4 | 94.0 |
| Q4_K_M | 89.44 GiB | 90.4 | 90.5 | 90.7 | 91.1 | 91.9 | 93.5 |
| UD-Q5_K_M | 81.83 GiB | 82.8 | 82.9 | 83.1 | 83.5 | 84.3 | 85.9 |
| Q5_K_L | 78.39 GiB | 79.4 | 79.5 | 79.7 | 80.1 | 80.9 | 82.5 |
| Q5_K_M | 78.22 GiB | 79.2 | 79.3 | 79.5 | 79.9 | 80.7 | 82.3 |
| UD-Q5_K_S | 76.98 GiB | 78.0 | 78.1 | 78.3 | 78.7 | 79.5 | 81.1 |
| Q5_K_S | 75.72 GiB | 76.7 | 76.8 | 77.0 | 77.4 | 78.2 | 79.8 |
| Q4_1 | 68.96 GiB | 70.0 | 70.1 | 70.3 | 70.7 | 71.4 | 73.0 |
| UD-Q4_K_M | 68.10 GiB | 69.1 | 69.2 | 69.4 | 69.8 | 70.6 | 72.2 |
| Q4_K_S | 64.36 GiB | 65.4 | 65.5 | 65.7 | 66.1 | 66.9 | 68.4 |
| UD-Q4_K_S | 63.88 GiB | 64.9 | 65.0 | 65.2 | 65.6 | 66.4 | 68.0 |
| Q4_0 | 62.44 GiB | 63.4 | 63.5 | 63.7 | 64.1 | 64.9 | 66.5 |
| IQ4_NL | 62.32 GiB | 63.3 | 63.4 | 63.6 | 64.0 | 64.8 | 66.4 |
| IQ4_XS | 58.98 GiB | 60.0 | 60.1 | 60.3 | 60.7 | 61.5 | 63.1 |
| UD-IQ4_NL | 54.71 GiB | 55.7 | 55.8 | 56.0 | 56.4 | 57.2 | 58.8 |
| UD-IQ4_XS | 53.61 GiB | 54.6 | 54.7 | 54.9 | 55.3 | 56.1 | 57.7 |
| IQ3_M | 52.60 GiB | 53.6 | 53.7 | 53.9 | 54.3 | 55.1 | 56.7 |
| Q3_K_L | 52.33 GiB | 53.3 | 53.4 | 53.6 | 54.0 | 54.8 | 56.4 |
| Q3_K_M | 50.33 GiB | 51.3 | 51.4 | 51.6 | 52.0 | 52.8 | 54.4 |
| UD-Q3_K_M | 50.31 GiB | 51.3 | 51.4 | 51.6 | 52.0 | 52.8 | 54.4 |
| IQ3_XS | 50.29 GiB | 51.3 | 51.4 | 51.6 | 52.0 | 52.8 | 54.4 |
| Q3_K_S | 48.03 GiB | 49.0 | 49.1 | 49.3 | 49.7 | 50.5 | 52.1 |
| IQ3_XXS | 46.07 GiB | 47.1 | 47.2 | 47.4 | 47.8 | 48.6 | 50.2 |
| UD-IQ3_S | 45.10 GiB | 46.1 | 46.2 | 46.4 | 46.8 | 47.6 | 49.2 |
| UD-IQ3_XXS | 41.24 GiB | 42.2 | 42.3 | 42.5 | 42.9 | 43.7 | 45.3 |
| Q2_K_L | 39.20 GiB | 40.2 | 40.3 | 40.5 | 40.9 | 41.7 | 43.3 |
| Q2_K | 38.92 GiB | 39.9 | 40.0 | 40.2 | 40.6 | 41.4 | 43.0 |
| IQ2_M | 37.19 GiB | 38.2 | 38.3 | 38.5 | 38.9 | 39.7 | 41.3 |
| UD-IQ2_M | 34.71 GiB | 35.7 | 35.8 | 36.0 | 36.4 | 37.2 | 38.8 |
| UD-IQ2_XXS | 34.64 GiB | 35.6 | 35.7 | 35.9 | 36.3 | 37.1 | 38.7 |
| IQ2_S | 33.77 GiB | 34.8 | 34.9 | 35.1 | 35.5 | 36.3 | 37.9 |
| UD-IQ1_M | 33.19 GiB | 34.2 | 34.3 | 34.5 | 34.9 | 35.7 | 37.3 |
| IQ2_XS | 33.12 GiB | 34.1 | 34.2 | 34.4 | 34.8 | 35.6 | 37.2 |
| UD-IQ1_S | 31.45 GiB | 32.4 | 32.5 | 32.7 | 33.1 | 33.9 | 35.5 |
| IQ2_XXS | 29.82 GiB | 30.8 | 30.9 | 31.1 | 31.5 | 32.3 | 33.9 |
| IQ1_M | 25.75 GiB | 26.7 | 26.8 | 27.0 | 27.4 | 28.2 | 29.8 |
| IQ1_S | 23.15 GiB | 24.1 | 24.2 | 24.4 | 24.8 | 25.6 | 27.2 |
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 512-token window rather than the full context.