Can I run Llama-4-Maverick-17B-128E-Instruct on a Apple M3 Pro?
Not at these settings. No indexed quantization of Llama-4-Maverick-17B-128E-Instruct fits Apple M3 Pro at any context we compute, with q8_0 KV. The smallest shipped quantization is 92.59 GiB in weights alone, against 12.56 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 | 747.4 | 747.8 | 748.6 | 750.2 | 753.4 | 759.8 |
| Q8_0 | 396.57 GiB | 397.5 | 397.9 | 398.7 | 400.3 | 403.5 | 409.9 |
| Q6_K | 306.19 GiB | 307.2 | 307.6 | 308.4 | 310.0 | 313.1 | 319.5 |
| UD-IQ2_M | 273.65 GiB | 274.6 | 275.0 | 275.8 | 277.4 | 280.6 | 287.0 |
| Q5_K_M | 264.93 GiB | 265.9 | 266.3 | 267.1 | 268.7 | 271.9 | 278.3 |
| Q5_K_S | 256.76 GiB | 257.7 | 258.1 | 258.9 | 260.5 | 263.7 | 270.1 |
| Q4_1 | 233.50 GiB | 234.5 | 234.9 | 235.7 | 237.3 | 240.4 | 246.8 |
| Q4_K_M | 226.09 GiB | 227.1 | 227.5 | 228.3 | 229.9 | 233.0 | 239.4 |
| Q4_K_S | 212.15 GiB | 213.1 | 213.5 | 214.3 | 215.9 | 219.1 | 225.5 |
| Q4_0 | 211.19 GiB | 212.2 | 212.6 | 213.4 | 215.0 | 218.1 | 224.5 |
| IQ4_NL | 210.26 GiB | 211.2 | 211.6 | 212.4 | 214.0 | 217.2 | 223.6 |
| UD-IQ4_XS | 205.52 GiB | 206.5 | 206.9 | 207.7 | 209.3 | 212.5 | 218.8 |
| IQ4_XS | 199.61 GiB | 200.6 | 201.0 | 201.8 | 203.4 | 206.6 | 212.9 |
| Q3_K_M | 177.95 GiB | 178.9 | 179.3 | 180.1 | 181.7 | 184.9 | 191.3 |
| Q3_K_S | 160.80 GiB | 161.8 | 162.2 | 163.0 | 164.6 | 167.7 | 174.1 |
| UD-IQ3_XXS | 157.69 GiB | 158.7 | 159.1 | 159.9 | 161.5 | 164.6 | 171.0 |
| Q2_K_L | 135.87 GiB | 136.8 | 137.2 | 138.0 | 139.6 | 142.8 | 149.2 |
| Q2_K | 135.64 GiB | 136.6 | 137.0 | 137.8 | 139.4 | 142.6 | 149.0 |
| UD-IQ2_XXS | 125.91 GiB | 126.9 | 127.3 | 128.1 | 129.7 | 132.9 | 139.2 |
| UD-IQ1_M | 118.78 GiB | 119.8 | 120.2 | 121.0 | 122.5 | 125.7 | 132.1 |
| UD-IQ1_S | 112.48 GiB | 113.5 | 113.8 | 114.6 | 116.2 | 119.4 | 125.8 |
| UD-TQ1_0 | 98.45 GiB | 99.4 | 99.8 | 100.6 | 102.2 | 105.4 | 111.8 |
| TQ1_0 | 92.59 GiB | 93.6 | 94.0 | 94.8 | 96.4 | 99.5 | 105.9 |
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 8,192-token window rather than the full context.