Can I run Mistral-Small-4-119B-2603 on a GeForce RTX 5090 D V2?
Not at these settings. No indexed quantization of Mistral-Small-4-119B-2603 fits GeForce RTX 5090 D V2 at any context we compute, with q8_0 KV. The smallest shipped quantization is 25.04 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 | 221.65 GiB | 222.5 | 222.6 | 222.7 | 222.9 | 223.2 | 224.0 |
| Q8_0 | 117.79 GiB | 118.7 | 118.7 | 118.8 | 119.0 | 119.4 | 120.1 |
| Q6_K_L | 95.99 GiB | 96.9 | 96.9 | 97.0 | 97.2 | 97.6 | 98.3 |
| UD-Q6_K | 92.60 GiB | 93.5 | 93.5 | 93.6 | 93.8 | 94.2 | 94.9 |
| Q6_K | 90.96 GiB | 91.8 | 91.9 | 92.0 | 92.2 | 92.5 | 93.3 |
| UD-Q5_K_M | 83.04 GiB | 83.9 | 84.0 | 84.1 | 84.2 | 84.6 | 85.4 |
| Q5_K_L | 79.51 GiB | 80.4 | 80.4 | 80.5 | 80.7 | 81.1 | 81.8 |
| Q5_K_M | 79.20 GiB | 80.1 | 80.1 | 80.2 | 80.4 | 80.8 | 81.5 |
| UD-Q5_K_S | 77.82 GiB | 78.7 | 78.7 | 78.8 | 79.0 | 79.4 | 80.1 |
| Q5_K_S | 76.57 GiB | 77.4 | 77.5 | 77.6 | 77.8 | 78.1 | 78.9 |
| UD-Q4_K_M | 68.70 GiB | 69.6 | 69.6 | 69.7 | 69.9 | 70.3 | 71.0 |
| Q4_K_L | 68.02 GiB | 68.9 | 68.9 | 69.0 | 69.2 | 69.6 | 70.3 |
| Q4_K_M | 67.65 GiB | 68.5 | 68.6 | 68.7 | 68.9 | 69.2 | 70.0 |
| Q4_K_S | 65.08 GiB | 66.0 | 66.0 | 66.1 | 66.3 | 66.7 | 67.4 |
| UD-Q4_K_S | 64.70 GiB | 65.6 | 65.6 | 65.7 | 65.9 | 66.3 | 67.0 |
| Q4_0 | 63.06 GiB | 63.9 | 64.0 | 64.1 | 64.3 | 64.6 | 65.4 |
| IQ4_NL | 62.83 GiB | 63.7 | 63.8 | 63.8 | 64.0 | 64.4 | 65.2 |
| IQ4_XS | 59.41 GiB | 60.3 | 60.3 | 60.4 | 60.6 | 61.0 | 61.7 |
| UD-IQ4_NL | 55.13 GiB | 56.0 | 56.1 | 56.2 | 56.3 | 56.7 | 57.5 |
| UD-IQ4_XS | 54.13 GiB | 55.0 | 55.1 | 55.2 | 55.3 | 55.7 | 56.5 |
| IQ3_M | 53.05 GiB | 53.9 | 54.0 | 54.1 | 54.3 | 54.6 | 55.4 |
| Q3_K_L | 52.99 GiB | 53.9 | 53.9 | 54.0 | 54.2 | 54.6 | 55.3 |
| Q3_K_M | 50.82 GiB | 51.7 | 51.7 | 51.8 | 52.0 | 52.4 | 53.1 |
| UD-Q3_K_M | 50.64 GiB | 51.5 | 51.6 | 51.7 | 51.8 | 52.2 | 53.0 |
| Q3_K_S | 48.42 GiB | 49.3 | 49.4 | 49.4 | 49.6 | 50.0 | 50.8 |
| IQ3_XS | 48.20 GiB | 49.1 | 49.1 | 49.2 | 49.4 | 49.8 | 50.5 |
| IQ3_XXS | 46.44 GiB | 47.3 | 47.4 | 47.5 | 47.6 | 48.0 | 48.8 |
| UD-Q3_K_S | 46.23 GiB | 47.1 | 47.2 | 47.3 | 47.4 | 47.8 | 48.6 |
| UD-IQ3_S | 41.36 GiB | 42.2 | 42.3 | 42.4 | 42.6 | 42.9 | 43.7 |
| UD-IQ3_XXS | 39.86 GiB | 40.7 | 40.8 | 40.9 | 41.1 | 41.4 | 42.2 |
| Q2_K_L | 39.67 GiB | 40.5 | 40.6 | 40.7 | 40.9 | 41.2 | 42.0 |
| Q2_K | 39.18 GiB | 40.1 | 40.1 | 40.2 | 40.4 | 40.8 | 41.5 |
| IQ2_M | 38.77 GiB | 39.6 | 39.7 | 39.8 | 40.0 | 40.3 | 41.1 |
| IQ2_S | 35.52 GiB | 36.4 | 36.4 | 36.5 | 36.7 | 37.1 | 37.8 |
| UD-IQ2_M | 34.99 GiB | 35.9 | 35.9 | 36.0 | 36.2 | 36.6 | 37.3 |
| IQ2_XS | 34.48 GiB | 35.4 | 35.4 | 35.5 | 35.7 | 36.1 | 36.8 |
| UD-IQ2_XXS | 32.53 GiB | 33.4 | 33.5 | 33.5 | 33.7 | 34.1 | 34.9 |
| IQ2_XXS | 31.35 GiB | 32.2 | 32.3 | 32.4 | 32.6 | 32.9 | 33.7 |
| UD-IQ1_M | 30.09 GiB | 31.0 | 31.0 | 31.1 | 31.3 | 31.7 | 32.4 |
| IQ1_M | 27.51 GiB | 28.4 | 28.4 | 28.5 | 28.7 | 29.1 | 29.8 |
| IQ1_S | 25.04 GiB | 25.9 | 26.0 | 26.1 | 26.2 | 26.6 | 27.4 |
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, this model uses latent attention and allocates no V cache at all, so any formula reading num_key_value_heads overstates its cache by more than an order of magnitude.