Can I run MiMo-V2.5 on a Apple M5 Max?
Not at these settings. No indexed quantization of MiMo-V2.5 fits Apple M5 Max at any context we compute, with q8_0 KV. The smallest shipped quantization is 60.09 GiB in weights alone, against 25.11 GiB usable. CPU offload can still run it, slowly.
Every quantization at every context
| Quant | Weights● | 4K◐ | 8K◐ | 16K◐ | 32K◐ | 64K◐ | 128K◐ |
|---|---|---|---|---|---|---|---|
| BF16 | 577.08 GiB | 577.9 | 578.2 | 578.7 | 579.7 | 581.7 | 585.7 |
| Q8_0 | 306.67 GiB | 307.5 | 307.8 | 308.3 | 309.3 | 311.2 | 315.2 |
| Q6_K | 249.18 GiB | 250.0 | 250.3 | 250.8 | 251.8 | 253.8 | 257.7 |
| UD-Q6_K | 239.34 GiB | 240.2 | 240.4 | 240.9 | 241.9 | 243.9 | 247.9 |
| UD-Q5_K_M | 214.37 GiB | 215.2 | 215.5 | 216.0 | 217.0 | 219.0 | 222.9 |
| Q5_K_M | 206.24 GiB | 207.1 | 207.3 | 207.8 | 208.8 | 210.8 | 214.8 |
| UD-Q5_K_S | 201.45 GiB | 202.3 | 202.5 | 203.0 | 204.0 | 206.0 | 210.0 |
| Q5_K_S | 199.54 GiB | 200.4 | 200.6 | 201.1 | 202.1 | 204.1 | 208.1 |
| Q4_1 | 180.99 GiB | 181.8 | 182.1 | 182.6 | 183.6 | 185.6 | 189.6 |
| UD-Q4_K_M | 177.81 GiB | 178.7 | 178.9 | 179.4 | 180.4 | 182.4 | 186.4 |
| Q4_K_L | 176.24 GiB | 177.1 | 177.3 | 177.8 | 178.8 | 180.8 | 184.8 |
| Q4_K_M | 175.80 GiB | 176.7 | 176.9 | 177.4 | 178.4 | 180.4 | 184.4 |
| Q4_K_S | 169.41 GiB | 170.3 | 170.5 | 171.0 | 172.0 | 174.0 | 178.0 |
| UD-Q4_K_S | 166.56 GiB | 167.4 | 167.7 | 168.1 | 169.1 | 171.1 | 175.1 |
| Q4_K | 165.75 GiB | 166.6 | 166.8 | 167.3 | 168.3 | 170.3 | 174.3 |
| Q4_0 | 163.68 GiB | 164.5 | 164.8 | 165.3 | 166.3 | 168.3 | 172.2 |
| IQ4_NL | 163.24 GiB | 164.1 | 164.3 | 164.8 | 165.8 | 167.8 | 171.8 |
| IQ4_XS | 154.32 GiB | 155.2 | 155.4 | 155.9 | 156.9 | 158.9 | 162.9 |
| UD-IQ4_NL | 142.12 GiB | 143.0 | 143.2 | 143.7 | 144.7 | 146.7 | 150.7 |
| UD-IQ4_XS | 139.18 GiB | 140.0 | 140.3 | 140.8 | 141.8 | 143.8 | 147.7 |
| IQ3_M | 138.15 GiB | 139.0 | 139.2 | 139.7 | 140.7 | 142.7 | 146.7 |
| Q3_K_L | 137.77 GiB | 138.6 | 138.9 | 139.4 | 140.4 | 142.4 | 146.3 |
| Q3_K_M | 132.47 GiB | 133.3 | 133.6 | 134.1 | 135.1 | 137.1 | 141.0 |
| IQ3_XS | 132.46 GiB | 133.3 | 133.6 | 134.1 | 135.0 | 137.0 | 141.0 |
| UD-Q3_K_M | 130.41 GiB | 131.3 | 131.5 | 132.0 | 133.0 | 135.0 | 139.0 |
| Q3_K_S | 125.83 GiB | 126.7 | 126.9 | 127.4 | 128.4 | 130.4 | 134.4 |
| IQ3_XXS | 121.31 GiB | 122.2 | 122.4 | 122.9 | 123.9 | 125.9 | 129.9 |
| UD-IQ3_XXS | 117.27 GiB | 118.1 | 118.4 | 118.9 | 119.9 | 121.9 | 125.8 |
| UD-IQ3_S | 106.98 GiB | 107.8 | 108.1 | 108.6 | 109.6 | 111.6 | 115.5 |
| IQ3_S | 106.20 GiB | 107.0 | 107.3 | 107.8 | 108.8 | 110.8 | 114.8 |
| Q2_K_L | 102.02 GiB | 102.9 | 103.1 | 103.6 | 104.6 | 106.6 | 110.6 |
| Q2_K | 101.46 GiB | 102.3 | 102.6 | 103.1 | 104.0 | 106.0 | 110.0 |
| IQ2_M | 97.30 GiB | 98.1 | 98.4 | 98.9 | 99.9 | 101.9 | 105.9 |
| UD-IQ2_M | 89.93 GiB | 90.8 | 91.0 | 91.5 | 92.5 | 94.5 | 98.5 |
| UD-IQ2_XXS | 89.85 GiB | 90.7 | 91.0 | 91.4 | 92.4 | 94.4 | 98.4 |
| IQ2_S | 88.18 GiB | 89.0 | 89.3 | 89.8 | 90.8 | 92.8 | 96.7 |
| IQ2_XS | 86.68 GiB | 87.5 | 87.8 | 88.3 | 89.3 | 91.3 | 95.2 |
| UD-IQ1_M | 86.17 GiB | 87.0 | 87.3 | 87.8 | 88.8 | 90.7 | 94.7 |
| IQ2_XXS | 79.83 GiB | 80.7 | 80.9 | 81.4 | 82.4 | 84.4 | 88.4 |
| IQ1_M | 67.01 GiB | 67.9 | 68.1 | 68.6 | 69.6 | 71.6 | 75.6 |
| IQ1_S | 60.09 GiB | 60.9 | 61.2 | 61.7 | 62.7 | 64.7 | 68.7 |
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 128-token window rather than the full context.