Can I run Llama-4-Maverick-17B-128E-Instruct on a GeForce RTX 5090 D?
Not at these settings. No indexed quantization of Llama-4-Maverick-17B-128E-Instruct fits GeForce RTX 5090 D at any context we compute, with q8_0 KV. The smallest shipped quantization is 92.59 GiB in weights alone, against 29.76 GiB usable. CPU offload can still run it, slowly.
Every quantization at every context
| Quant | Weights● | 4K◐ | 8K◐ | 16K◐ | 32K◐ | 64K◐ | 128K◐ |
|---|---|---|---|---|---|---|---|
| BF16 | 746.43 GiB | 747.7 | 748.0 | 748.8 | 750.4 | 753.6 | 760.0 |
| Q8_0 | 396.57 GiB | 397.8 | 398.2 | 399.0 | 400.6 | 403.8 | 410.1 |
| Q6_K | 306.19 GiB | 307.4 | 307.8 | 308.6 | 310.2 | 313.4 | 319.8 |
| UD-IQ2_M | 273.65 GiB | 274.9 | 275.3 | 276.1 | 277.7 | 280.8 | 287.2 |
| Q5_K_M | 264.93 GiB | 266.2 | 266.6 | 267.4 | 268.9 | 272.1 | 278.5 |
| Q5_K_S | 256.76 GiB | 258.0 | 258.4 | 259.2 | 260.8 | 264.0 | 270.3 |
| Q4_1 | 233.50 GiB | 234.7 | 235.1 | 235.9 | 237.5 | 240.7 | 247.1 |
| Q4_K_M | 226.09 GiB | 227.3 | 227.7 | 228.5 | 230.1 | 233.3 | 239.7 |
| Q4_K_S | 212.15 GiB | 213.4 | 213.8 | 214.6 | 216.2 | 219.4 | 225.7 |
| Q4_0 | 211.19 GiB | 212.4 | 212.8 | 213.6 | 215.2 | 218.4 | 224.8 |
| IQ4_NL | 210.26 GiB | 211.5 | 211.9 | 212.7 | 214.3 | 217.5 | 223.8 |
| UD-IQ4_XS | 205.52 GiB | 206.7 | 207.1 | 207.9 | 209.5 | 212.7 | 219.1 |
| IQ4_XS | 199.61 GiB | 200.8 | 201.2 | 202.0 | 203.6 | 206.8 | 213.2 |
| Q3_K_M | 177.95 GiB | 179.2 | 179.6 | 180.4 | 182.0 | 185.1 | 191.5 |
| Q3_K_S | 160.80 GiB | 162.0 | 162.4 | 163.2 | 164.8 | 168.0 | 174.4 |
| UD-IQ3_XXS | 157.69 GiB | 158.9 | 159.3 | 160.1 | 161.7 | 164.9 | 171.3 |
| Q2_K_L | 135.87 GiB | 137.1 | 137.5 | 138.3 | 139.9 | 143.1 | 149.4 |
| Q2_K | 135.64 GiB | 136.9 | 137.3 | 138.1 | 139.7 | 142.8 | 149.2 |
| UD-IQ2_XXS | 125.91 GiB | 127.1 | 127.5 | 128.3 | 129.9 | 133.1 | 139.5 |
| UD-IQ1_M | 118.78 GiB | 120.0 | 120.4 | 121.2 | 122.8 | 126.0 | 132.4 |
| UD-IQ1_S | 112.48 GiB | 113.7 | 114.1 | 114.9 | 116.5 | 119.7 | 126.1 |
| UD-TQ1_0 | 98.45 GiB | 99.7 | 100.1 | 100.9 | 102.5 | 105.7 | 112.0 |
| TQ1_0 | 92.59 GiB | 93.8 | 94.2 | 95.0 | 96.6 | 99.8 | 106.2 |
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 8,192-token window rather than the full context.