Can I run MiMo-V2.5 on a Radeon RX 6500 XT?
Not at these settings. No indexed quantization of MiMo-V2.5 fits Radeon RX 6500 XT at any context we compute, with q8_0 KV. The smallest shipped quantization is 60.09 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 | 577.08 GiB | 578.3 | 578.5 | 579.0 | 580.0 | 582.0 | 586.0 |
| Q8_0 | 306.67 GiB | 307.9 | 308.1 | 308.6 | 309.6 | 311.6 | 315.6 |
| Q6_K | 249.18 GiB | 250.4 | 250.6 | 251.1 | 252.1 | 254.1 | 258.1 |
| UD-Q6_K | 239.34 GiB | 240.5 | 240.8 | 241.3 | 242.3 | 244.3 | 248.3 |
| UD-Q5_K_M | 214.37 GiB | 215.6 | 215.8 | 216.3 | 217.3 | 219.3 | 223.3 |
| Q5_K_M | 206.24 GiB | 207.4 | 207.7 | 208.2 | 209.2 | 211.2 | 215.2 |
| UD-Q5_K_S | 201.45 GiB | 202.6 | 202.9 | 203.4 | 204.4 | 206.4 | 210.4 |
| Q5_K_S | 199.54 GiB | 200.7 | 201.0 | 201.5 | 202.5 | 204.5 | 208.5 |
| Q4_1 | 180.99 GiB | 182.2 | 182.4 | 182.9 | 183.9 | 185.9 | 189.9 |
| UD-Q4_K_M | 177.81 GiB | 179.0 | 179.3 | 179.7 | 180.7 | 182.7 | 186.7 |
| Q4_K_L | 176.24 GiB | 177.4 | 177.7 | 178.2 | 179.2 | 181.2 | 185.2 |
| Q4_K_M | 175.80 GiB | 177.0 | 177.3 | 177.7 | 178.7 | 180.7 | 184.7 |
| Q4_K_S | 169.41 GiB | 170.6 | 170.9 | 171.4 | 172.4 | 174.3 | 178.3 |
| UD-Q4_K_S | 166.56 GiB | 167.8 | 168.0 | 168.5 | 169.5 | 171.5 | 175.5 |
| Q4_K | 165.75 GiB | 167.0 | 167.2 | 167.7 | 168.7 | 170.7 | 174.7 |
| Q4_0 | 163.68 GiB | 164.9 | 165.1 | 165.6 | 166.6 | 168.6 | 172.6 |
| IQ4_NL | 163.24 GiB | 164.4 | 164.7 | 165.2 | 166.2 | 168.2 | 172.2 |
| IQ4_XS | 154.32 GiB | 155.5 | 155.8 | 156.3 | 157.3 | 159.3 | 163.2 |
| UD-IQ4_NL | 142.12 GiB | 143.3 | 143.6 | 144.1 | 145.1 | 147.1 | 151.0 |
| UD-IQ4_XS | 139.18 GiB | 140.4 | 140.6 | 141.1 | 142.1 | 144.1 | 148.1 |
| IQ3_M | 138.15 GiB | 139.3 | 139.6 | 140.1 | 141.1 | 143.1 | 147.1 |
| Q3_K_L | 137.77 GiB | 139.0 | 139.2 | 139.7 | 140.7 | 142.7 | 146.7 |
| Q3_K_M | 132.47 GiB | 133.7 | 133.9 | 134.4 | 135.4 | 137.4 | 141.4 |
| IQ3_XS | 132.46 GiB | 133.7 | 133.9 | 134.4 | 135.4 | 137.4 | 141.4 |
| UD-Q3_K_M | 130.41 GiB | 131.6 | 131.9 | 132.4 | 133.4 | 135.3 | 139.3 |
| Q3_K_S | 125.83 GiB | 127.0 | 127.3 | 127.8 | 128.8 | 130.8 | 134.7 |
| IQ3_XXS | 121.31 GiB | 122.5 | 122.8 | 123.3 | 124.2 | 126.2 | 130.2 |
| UD-IQ3_XXS | 117.27 GiB | 118.5 | 118.7 | 119.2 | 120.2 | 122.2 | 126.2 |
| UD-IQ3_S | 106.98 GiB | 108.2 | 108.4 | 108.9 | 109.9 | 111.9 | 115.9 |
| IQ3_S | 106.20 GiB | 107.4 | 107.6 | 108.1 | 109.1 | 111.1 | 115.1 |
| Q2_K_L | 102.02 GiB | 103.2 | 103.5 | 104.0 | 105.0 | 107.0 | 110.9 |
| Q2_K | 101.46 GiB | 102.7 | 102.9 | 103.4 | 104.4 | 106.4 | 110.4 |
| IQ2_M | 97.30 GiB | 98.5 | 98.7 | 99.2 | 100.2 | 102.2 | 106.2 |
| UD-IQ2_M | 89.93 GiB | 91.1 | 91.4 | 91.9 | 92.9 | 94.9 | 98.8 |
| UD-IQ2_XXS | 89.85 GiB | 91.1 | 91.3 | 91.8 | 92.8 | 94.8 | 98.8 |
| IQ2_S | 88.18 GiB | 89.4 | 89.6 | 90.1 | 91.1 | 93.1 | 97.1 |
| IQ2_XS | 86.68 GiB | 87.9 | 88.1 | 88.6 | 89.6 | 91.6 | 95.6 |
| UD-IQ1_M | 86.17 GiB | 87.4 | 87.6 | 88.1 | 89.1 | 91.1 | 95.1 |
| IQ2_XXS | 79.83 GiB | 81.0 | 81.3 | 81.8 | 82.8 | 84.8 | 88.7 |
| IQ1_M | 67.01 GiB | 68.2 | 68.5 | 69.0 | 69.9 | 71.9 | 75.9 |
| IQ1_S | 60.09 GiB | 61.3 | 61.5 | 62.0 | 63.0 | 65.0 | 69.0 |
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.