Can I run Llama-4-Maverick-17B-128E-Instruct on a Radeon RX 6500 XT?
Not at these settings. No indexed quantization of Llama-4-Maverick-17B-128E-Instruct fits Radeon RX 6500 XT at any context we compute, with q4_0 KV. The smallest shipped quantization is 92.59 GiB in weights alone, against 3.72 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 | 747.6 | 747.8 | 748.2 | 749.0 | 750.7 | 754.1 |
| Q8_0 | 396.57 GiB | 397.7 | 397.9 | 398.3 | 399.2 | 400.9 | 404.2 |
| Q6_K | 306.19 GiB | 307.3 | 307.5 | 308.0 | 308.8 | 310.5 | 313.9 |
| UD-IQ2_M | 273.65 GiB | 274.8 | 275.0 | 275.4 | 276.3 | 277.9 | 281.3 |
| Q5_K_M | 264.93 GiB | 266.1 | 266.3 | 266.7 | 267.5 | 269.2 | 272.6 |
| Q5_K_S | 256.76 GiB | 257.9 | 258.1 | 258.5 | 259.4 | 261.1 | 264.4 |
| Q4_1 | 233.50 GiB | 234.6 | 234.8 | 235.3 | 236.1 | 237.8 | 241.2 |
| Q4_K_M | 226.09 GiB | 227.2 | 227.4 | 227.9 | 228.7 | 230.4 | 233.8 |
| Q4_K_S | 212.15 GiB | 213.3 | 213.5 | 213.9 | 214.8 | 216.5 | 219.8 |
| Q4_0 | 211.19 GiB | 212.3 | 212.5 | 213.0 | 213.8 | 215.5 | 218.9 |
| IQ4_NL | 210.26 GiB | 211.4 | 211.6 | 212.0 | 212.9 | 214.6 | 217.9 |
| UD-IQ4_XS | 205.52 GiB | 206.7 | 206.9 | 207.3 | 208.1 | 209.8 | 213.2 |
| IQ4_XS | 199.61 GiB | 200.7 | 201.0 | 201.4 | 202.2 | 203.9 | 207.3 |
| Q3_K_M | 177.95 GiB | 179.1 | 179.3 | 179.7 | 180.6 | 182.2 | 185.6 |
| Q3_K_S | 160.80 GiB | 161.9 | 162.1 | 162.6 | 163.4 | 165.1 | 168.5 |
| UD-IQ3_XXS | 157.69 GiB | 158.8 | 159.0 | 159.5 | 160.3 | 162.0 | 165.4 |
| Q2_K_L | 135.87 GiB | 137.0 | 137.2 | 137.6 | 138.5 | 140.2 | 143.5 |
| Q2_K | 135.64 GiB | 136.8 | 137.0 | 137.4 | 138.3 | 139.9 | 143.3 |
| UD-IQ2_XXS | 125.91 GiB | 127.0 | 127.3 | 127.7 | 128.5 | 130.2 | 133.6 |
| UD-IQ1_M | 118.78 GiB | 119.9 | 120.1 | 120.6 | 121.4 | 123.1 | 126.5 |
| UD-IQ1_S | 112.48 GiB | 113.6 | 113.8 | 114.2 | 115.1 | 116.8 | 120.2 |
| UD-TQ1_0 | 98.45 GiB | 99.6 | 99.8 | 100.2 | 101.1 | 102.8 | 106.1 |
| TQ1_0 | 92.59 GiB | 93.7 | 93.9 | 94.4 | 95.2 | 96.9 | 100.3 |
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.