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 q4_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 | 578.2 | 578.3 | 578.6 | 579.1 | 580.1 | 582.3 |
| Q8_0 | 306.67 GiB | 307.7 | 307.9 | 308.1 | 308.7 | 309.7 | 311.8 |
| Q6_K | 249.18 GiB | 250.3 | 250.4 | 250.7 | 251.2 | 252.2 | 254.3 |
| UD-Q6_K | 239.34 GiB | 240.4 | 240.5 | 240.8 | 241.3 | 242.4 | 244.5 |
| UD-Q5_K_M | 214.37 GiB | 215.4 | 215.6 | 215.8 | 216.4 | 217.4 | 219.5 |
| Q5_K_M | 206.24 GiB | 207.3 | 207.5 | 207.7 | 208.2 | 209.3 | 211.4 |
| UD-Q5_K_S | 201.45 GiB | 202.5 | 202.7 | 202.9 | 203.4 | 204.5 | 206.6 |
| Q5_K_S | 199.54 GiB | 200.6 | 200.8 | 201.0 | 201.5 | 202.6 | 204.7 |
| Q4_1 | 180.99 GiB | 182.1 | 182.2 | 182.5 | 183.0 | 184.0 | 186.2 |
| UD-Q4_K_M | 177.81 GiB | 178.9 | 179.0 | 179.3 | 179.8 | 180.9 | 183.0 |
| Q4_K_L | 176.24 GiB | 177.3 | 177.4 | 177.7 | 178.2 | 179.3 | 181.4 |
| Q4_K_M | 175.80 GiB | 176.9 | 177.0 | 177.3 | 177.8 | 178.9 | 181.0 |
| Q4_K_S | 169.41 GiB | 170.5 | 170.6 | 170.9 | 171.4 | 172.5 | 174.6 |
| UD-Q4_K_S | 166.56 GiB | 167.6 | 167.8 | 168.0 | 168.6 | 169.6 | 171.7 |
| Q4_K | 165.75 GiB | 166.8 | 167.0 | 167.2 | 167.8 | 168.8 | 170.9 |
| Q4_0 | 163.68 GiB | 164.8 | 164.9 | 165.2 | 165.7 | 166.7 | 168.8 |
| IQ4_NL | 163.24 GiB | 164.3 | 164.5 | 164.7 | 165.2 | 166.3 | 168.4 |
| IQ4_XS | 154.32 GiB | 155.4 | 155.5 | 155.8 | 156.3 | 157.4 | 159.5 |
| UD-IQ4_NL | 142.12 GiB | 143.2 | 143.3 | 143.6 | 144.1 | 145.2 | 147.3 |
| UD-IQ4_XS | 139.18 GiB | 140.3 | 140.4 | 140.7 | 141.2 | 142.2 | 144.3 |
| IQ3_M | 138.15 GiB | 139.2 | 139.4 | 139.6 | 140.2 | 141.2 | 143.3 |
| Q3_K_L | 137.77 GiB | 138.9 | 139.0 | 139.2 | 139.8 | 140.8 | 142.9 |
| Q3_K_M | 132.47 GiB | 133.6 | 133.7 | 133.9 | 134.5 | 135.5 | 137.6 |
| IQ3_XS | 132.46 GiB | 133.5 | 133.7 | 133.9 | 134.5 | 135.5 | 137.6 |
| UD-Q3_K_M | 130.41 GiB | 131.5 | 131.6 | 131.9 | 132.4 | 133.5 | 135.6 |
| Q3_K_S | 125.83 GiB | 126.9 | 127.0 | 127.3 | 127.8 | 128.9 | 131.0 |
| IQ3_XXS | 121.31 GiB | 122.4 | 122.5 | 122.8 | 123.3 | 124.4 | 126.5 |
| UD-IQ3_XXS | 117.27 GiB | 118.4 | 118.5 | 118.7 | 119.3 | 120.3 | 122.4 |
| UD-IQ3_S | 106.98 GiB | 108.1 | 108.2 | 108.5 | 109.0 | 110.0 | 112.1 |
| IQ3_S | 106.20 GiB | 107.3 | 107.4 | 107.7 | 108.2 | 109.3 | 111.4 |
| Q2_K_L | 102.02 GiB | 103.1 | 103.2 | 103.5 | 104.0 | 105.1 | 107.2 |
| Q2_K | 101.46 GiB | 102.5 | 102.7 | 102.9 | 103.5 | 104.5 | 106.6 |
| IQ2_M | 97.30 GiB | 98.4 | 98.5 | 98.8 | 99.3 | 100.4 | 102.5 |
| UD-IQ2_M | 89.93 GiB | 91.0 | 91.1 | 91.4 | 91.9 | 93.0 | 95.1 |
| UD-IQ2_XXS | 89.85 GiB | 90.9 | 91.1 | 91.3 | 91.9 | 92.9 | 95.0 |
| IQ2_S | 88.18 GiB | 89.3 | 89.4 | 89.7 | 90.2 | 91.2 | 93.3 |
| IQ2_XS | 86.68 GiB | 87.8 | 87.9 | 88.2 | 88.7 | 89.7 | 91.8 |
| UD-IQ1_M | 86.17 GiB | 87.2 | 87.4 | 87.6 | 88.2 | 89.2 | 91.3 |
| IQ2_XXS | 79.83 GiB | 80.9 | 81.0 | 81.3 | 81.8 | 82.9 | 85.0 |
| IQ1_M | 67.01 GiB | 68.1 | 68.2 | 68.5 | 69.0 | 70.1 | 72.2 |
| IQ1_S | 60.09 GiB | 61.2 | 61.3 | 61.6 | 62.1 | 63.1 | 65.3 |
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 128-token window rather than the full context.