Can I run Llama-4-Maverick-17B-128E-Instruct on a GeForce RTX 3050?
Not at these settings. No indexed quantization of Llama-4-Maverick-17B-128E-Instruct fits GeForce RTX 3050 at any context we compute, with q4_0 KV. The smallest shipped quantization is 92.59 GiB in weights alone, against 5.58 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 | 747.5 | 747.7 | 748.1 | 748.9 | 750.6 | 754.0 |
| Q8_0 | 396.57 GiB | 397.6 | 397.8 | 398.2 | 399.1 | 400.8 | 404.1 |
| Q6_K | 306.19 GiB | 307.2 | 307.4 | 307.9 | 308.7 | 310.4 | 313.8 |
| UD-IQ2_M | 273.65 GiB | 274.7 | 274.9 | 275.3 | 276.2 | 277.8 | 281.2 |
| Q5_K_M | 264.93 GiB | 266.0 | 266.2 | 266.6 | 267.4 | 269.1 | 272.5 |
| Q5_K_S | 256.76 GiB | 257.8 | 258.0 | 258.4 | 259.3 | 261.0 | 264.3 |
| Q4_1 | 233.50 GiB | 234.5 | 234.7 | 235.2 | 236.0 | 237.7 | 241.1 |
| Q4_K_M | 226.09 GiB | 227.1 | 227.3 | 227.8 | 228.6 | 230.3 | 233.7 |
| Q4_K_S | 212.15 GiB | 213.2 | 213.4 | 213.8 | 214.7 | 216.4 | 219.7 |
| Q4_0 | 211.19 GiB | 212.2 | 212.4 | 212.9 | 213.7 | 215.4 | 218.8 |
| IQ4_NL | 210.26 GiB | 211.3 | 211.5 | 211.9 | 212.8 | 214.5 | 217.8 |
| UD-IQ4_XS | 205.52 GiB | 206.6 | 206.8 | 207.2 | 208.0 | 209.7 | 213.1 |
| IQ4_XS | 199.61 GiB | 200.6 | 200.9 | 201.3 | 202.1 | 203.8 | 207.2 |
| Q3_K_M | 177.95 GiB | 179.0 | 179.2 | 179.6 | 180.5 | 182.1 | 185.5 |
| Q3_K_S | 160.80 GiB | 161.8 | 162.0 | 162.5 | 163.3 | 165.0 | 168.4 |
| UD-IQ3_XXS | 157.69 GiB | 158.7 | 158.9 | 159.4 | 160.2 | 161.9 | 165.3 |
| Q2_K_L | 135.87 GiB | 136.9 | 137.1 | 137.5 | 138.4 | 140.1 | 143.4 |
| Q2_K | 135.64 GiB | 136.7 | 136.9 | 137.3 | 138.2 | 139.8 | 143.2 |
| UD-IQ2_XXS | 125.91 GiB | 126.9 | 127.2 | 127.6 | 128.4 | 130.1 | 133.5 |
| UD-IQ1_M | 118.78 GiB | 119.8 | 120.0 | 120.5 | 121.3 | 123.0 | 126.4 |
| UD-IQ1_S | 112.48 GiB | 113.5 | 113.7 | 114.1 | 115.0 | 116.7 | 120.1 |
| UD-TQ1_0 | 98.45 GiB | 99.5 | 99.7 | 100.1 | 101.0 | 102.7 | 106.0 |
| TQ1_0 | 92.59 GiB | 93.6 | 93.8 | 94.3 | 95.1 | 96.8 | 100.2 |
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.