Can I run Laguna-S-2.1 on a RTX 4000 SFF Ada Generation?
Not at these settings. No indexed quantization of Laguna-S-2.1 fits RTX 4000 SFF Ada Generation at any context we compute, with q4_0 KV. The smallest shipped quantization is 23.15 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 | 219.05 GiB | 220.2 | 220.2 | 220.3 | 220.5 | 221.0 | 221.8 |
| F16 | 219.05 GiB | 220.2 | 220.2 | 220.3 | 220.5 | 221.0 | 221.8 |
| Q8_0 | 119.91 GiB | 121.0 | 121.1 | 121.2 | 121.4 | 121.8 | 122.7 |
| Q6_K_L | 94.83 GiB | 96.0 | 96.0 | 96.1 | 96.3 | 96.7 | 97.6 |
| UD-Q6_K | 91.19 GiB | 92.3 | 92.4 | 92.5 | 92.7 | 93.1 | 93.9 |
| Q6_K | 89.93 GiB | 91.0 | 91.1 | 91.2 | 91.4 | 91.8 | 92.7 |
| Q4_K_M | 89.44 GiB | 90.6 | 90.6 | 90.7 | 90.9 | 91.3 | 92.2 |
| UD-Q5_K_M | 81.83 GiB | 82.9 | 83.0 | 83.1 | 83.3 | 83.7 | 84.6 |
| Q5_K_L | 78.39 GiB | 79.5 | 79.6 | 79.7 | 79.9 | 80.3 | 81.1 |
| Q5_K_M | 78.22 GiB | 79.3 | 79.4 | 79.5 | 79.7 | 80.1 | 81.0 |
| UD-Q5_K_S | 76.98 GiB | 78.1 | 78.1 | 78.3 | 78.5 | 78.9 | 79.7 |
| Q5_K_S | 75.72 GiB | 76.8 | 76.9 | 77.0 | 77.2 | 77.6 | 78.5 |
| Q4_1 | 68.96 GiB | 70.1 | 70.1 | 70.2 | 70.4 | 70.9 | 71.7 |
| UD-Q4_K_M | 68.10 GiB | 69.2 | 69.3 | 69.4 | 69.6 | 70.0 | 70.8 |
| Q4_K_S | 64.36 GiB | 65.5 | 65.5 | 65.6 | 65.8 | 66.3 | 67.1 |
| UD-Q4_K_S | 63.88 GiB | 65.0 | 65.0 | 65.2 | 65.4 | 65.8 | 66.6 |
| Q4_0 | 62.44 GiB | 63.6 | 63.6 | 63.7 | 63.9 | 64.3 | 65.2 |
| IQ4_NL | 62.32 GiB | 63.4 | 63.5 | 63.6 | 63.8 | 64.2 | 65.1 |
| IQ4_XS | 58.98 GiB | 60.1 | 60.1 | 60.3 | 60.5 | 60.9 | 61.7 |
| UD-IQ4_NL | 54.71 GiB | 55.8 | 55.9 | 56.0 | 56.2 | 56.6 | 57.5 |
| UD-IQ4_XS | 53.61 GiB | 54.7 | 54.8 | 54.9 | 55.1 | 55.5 | 56.4 |
| IQ3_M | 52.60 GiB | 53.7 | 53.8 | 53.9 | 54.1 | 54.5 | 55.3 |
| Q3_K_L | 52.33 GiB | 53.4 | 53.5 | 53.6 | 53.8 | 54.2 | 55.1 |
| Q3_K_M | 50.33 GiB | 51.4 | 51.5 | 51.6 | 51.8 | 52.2 | 53.1 |
| UD-Q3_K_M | 50.31 GiB | 51.4 | 51.5 | 51.6 | 51.8 | 52.2 | 53.1 |
| IQ3_XS | 50.29 GiB | 51.4 | 51.5 | 51.6 | 51.8 | 52.2 | 53.0 |
| Q3_K_S | 48.03 GiB | 49.1 | 49.2 | 49.3 | 49.5 | 49.9 | 50.8 |
| IQ3_XXS | 46.07 GiB | 47.2 | 47.2 | 47.3 | 47.6 | 48.0 | 48.8 |
| UD-IQ3_S | 45.10 GiB | 46.2 | 46.3 | 46.4 | 46.6 | 47.0 | 47.9 |
| UD-IQ3_XXS | 41.24 GiB | 42.4 | 42.4 | 42.5 | 42.7 | 43.1 | 44.0 |
| Q2_K_L | 39.20 GiB | 40.3 | 40.4 | 40.5 | 40.7 | 41.1 | 41.9 |
| Q2_K | 38.92 GiB | 40.0 | 40.1 | 40.2 | 40.4 | 40.8 | 41.7 |
| IQ2_M | 37.19 GiB | 38.3 | 38.4 | 38.5 | 38.7 | 39.1 | 39.9 |
| UD-IQ2_M | 34.71 GiB | 35.8 | 35.9 | 36.0 | 36.2 | 36.6 | 37.5 |
| UD-IQ2_XXS | 34.64 GiB | 35.8 | 35.8 | 35.9 | 36.1 | 36.5 | 37.4 |
| IQ2_S | 33.77 GiB | 34.9 | 34.9 | 35.0 | 35.3 | 35.7 | 36.5 |
| UD-IQ1_M | 33.19 GiB | 34.3 | 34.4 | 34.5 | 34.7 | 35.1 | 35.9 |
| IQ2_XS | 33.12 GiB | 34.2 | 34.3 | 34.4 | 34.6 | 35.0 | 35.9 |
| UD-IQ1_S | 31.45 GiB | 32.6 | 32.6 | 32.7 | 32.9 | 33.4 | 34.2 |
| IQ2_XXS | 29.82 GiB | 30.9 | 31.0 | 31.1 | 31.3 | 31.7 | 32.6 |
| IQ1_M | 25.75 GiB | 26.9 | 26.9 | 27.0 | 27.2 | 27.7 | 28.5 |
| IQ1_S | 23.15 GiB | 24.3 | 24.3 | 24.4 | 24.6 | 25.1 | 25.9 |
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.