Can I run MiMo-V2.5 on a GeForce RTX 5070?
Not at these settings. No indexed quantization of MiMo-V2.5 fits GeForce RTX 5070 at any context we compute, with q8_0 KV. The smallest shipped quantization is 60.09 GiB in weights alone, against 11.16 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 | 578.4 | 578.9 | 579.9 | 581.9 | 585.9 |
| Q8_0 | 306.67 GiB | 307.8 | 308.0 | 308.5 | 309.5 | 311.5 | 315.5 |
| Q6_K | 249.18 GiB | 250.3 | 250.5 | 251.0 | 252.0 | 254.0 | 258.0 |
| UD-Q6_K | 239.34 GiB | 240.4 | 240.7 | 241.2 | 242.2 | 244.2 | 248.2 |
| UD-Q5_K_M | 214.37 GiB | 215.5 | 215.7 | 216.2 | 217.2 | 219.2 | 223.2 |
| Q5_K_M | 206.24 GiB | 207.3 | 207.6 | 208.1 | 209.1 | 211.1 | 215.1 |
| UD-Q5_K_S | 201.45 GiB | 202.5 | 202.8 | 203.3 | 204.3 | 206.3 | 210.3 |
| Q5_K_S | 199.54 GiB | 200.6 | 200.9 | 201.4 | 202.4 | 204.4 | 208.4 |
| Q4_1 | 180.99 GiB | 182.1 | 182.3 | 182.8 | 183.8 | 185.8 | 189.8 |
| UD-Q4_K_M | 177.81 GiB | 178.9 | 179.2 | 179.6 | 180.6 | 182.6 | 186.6 |
| Q4_K_L | 176.24 GiB | 177.3 | 177.6 | 178.1 | 179.1 | 181.1 | 185.1 |
| Q4_K_M | 175.80 GiB | 176.9 | 177.2 | 177.6 | 178.6 | 180.6 | 184.6 |
| Q4_K_S | 169.41 GiB | 170.5 | 170.8 | 171.3 | 172.3 | 174.2 | 178.2 |
| UD-Q4_K_S | 166.56 GiB | 167.7 | 167.9 | 168.4 | 169.4 | 171.4 | 175.4 |
| Q4_K | 165.75 GiB | 166.9 | 167.1 | 167.6 | 168.6 | 170.6 | 174.6 |
| Q4_0 | 163.68 GiB | 164.8 | 165.0 | 165.5 | 166.5 | 168.5 | 172.5 |
| IQ4_NL | 163.24 GiB | 164.3 | 164.6 | 165.1 | 166.1 | 168.1 | 172.1 |
| IQ4_XS | 154.32 GiB | 155.4 | 155.7 | 156.2 | 157.2 | 159.2 | 163.1 |
| UD-IQ4_NL | 142.12 GiB | 143.2 | 143.5 | 144.0 | 145.0 | 147.0 | 150.9 |
| UD-IQ4_XS | 139.18 GiB | 140.3 | 140.5 | 141.0 | 142.0 | 144.0 | 148.0 |
| IQ3_M | 138.15 GiB | 139.2 | 139.5 | 140.0 | 141.0 | 143.0 | 147.0 |
| Q3_K_L | 137.77 GiB | 138.9 | 139.1 | 139.6 | 140.6 | 142.6 | 146.6 |
| Q3_K_M | 132.47 GiB | 133.6 | 133.8 | 134.3 | 135.3 | 137.3 | 141.3 |
| IQ3_XS | 132.46 GiB | 133.6 | 133.8 | 134.3 | 135.3 | 137.3 | 141.3 |
| UD-Q3_K_M | 130.41 GiB | 131.5 | 131.8 | 132.3 | 133.3 | 135.2 | 139.2 |
| Q3_K_S | 125.83 GiB | 126.9 | 127.2 | 127.7 | 128.7 | 130.7 | 134.6 |
| IQ3_XXS | 121.31 GiB | 122.4 | 122.7 | 123.2 | 124.1 | 126.1 | 130.1 |
| UD-IQ3_XXS | 117.27 GiB | 118.4 | 118.6 | 119.1 | 120.1 | 122.1 | 126.1 |
| UD-IQ3_S | 106.98 GiB | 108.1 | 108.3 | 108.8 | 109.8 | 111.8 | 115.8 |
| IQ3_S | 106.20 GiB | 107.3 | 107.5 | 108.0 | 109.0 | 111.0 | 115.0 |
| Q2_K_L | 102.02 GiB | 103.1 | 103.4 | 103.9 | 104.9 | 106.9 | 110.8 |
| Q2_K | 101.46 GiB | 102.6 | 102.8 | 103.3 | 104.3 | 106.3 | 110.3 |
| IQ2_M | 97.30 GiB | 98.4 | 98.6 | 99.1 | 100.1 | 102.1 | 106.1 |
| UD-IQ2_M | 89.93 GiB | 91.0 | 91.3 | 91.8 | 92.8 | 94.8 | 98.7 |
| UD-IQ2_XXS | 89.85 GiB | 91.0 | 91.2 | 91.7 | 92.7 | 94.7 | 98.7 |
| IQ2_S | 88.18 GiB | 89.3 | 89.5 | 90.0 | 91.0 | 93.0 | 97.0 |
| IQ2_XS | 86.68 GiB | 87.8 | 88.0 | 88.5 | 89.5 | 91.5 | 95.5 |
| UD-IQ1_M | 86.17 GiB | 87.3 | 87.5 | 88.0 | 89.0 | 91.0 | 95.0 |
| IQ2_XXS | 79.83 GiB | 80.9 | 81.2 | 81.7 | 82.7 | 84.7 | 88.6 |
| IQ1_M | 67.01 GiB | 68.1 | 68.4 | 68.9 | 69.8 | 71.8 | 75.8 |
| IQ1_S | 60.09 GiB | 61.2 | 61.4 | 61.9 | 62.9 | 64.9 | 68.9 |
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 128-token window rather than the full context.