Can I run MiMo-V2.5-Pro on a GeForce RTX 3050?
Not at these settings. No indexed quantization of MiMo-V2.5-Pro fits GeForce RTX 3050 at any context we compute, with q8_0 KV. The smallest shipped quantization is 197.49 GiB in weights alone, against 5.58 GiB usable. CPU offload can still run it, slowly.
Every quantization at every context
| Quant | Weights● | 4K◐ | 8K◐ | 16K◐ | 32K◐ | 64K◐ | 128K◐ |
|---|---|---|---|---|---|---|---|
| BF16 | 1906.14 GiB | 1907.7 | 1908.5 | 1909.9 | 1912.8 | 1918.6 | 1930.2 |
| Q8_0 | 1012.92 GiB | 1014.5 | 1015.2 | 1016.7 | 1019.6 | 1025.4 | 1037.0 |
| UD-Q6_K | 788.42 GiB | 790.0 | 790.7 | 792.2 | 795.1 | 800.9 | 812.5 |
| UD-Q5_K_M | 705.94 GiB | 707.5 | 708.3 | 709.7 | 712.6 | 718.4 | 730.0 |
| Q5_K_M | 704.85 GiB | 706.4 | 707.2 | 708.6 | 711.5 | 717.3 | 729.0 |
| UD-Q5_K_S | 664.21 GiB | 665.8 | 666.5 | 668.0 | 670.9 | 676.7 | 688.3 |
| Q4_1 | 596.26 GiB | 597.9 | 598.6 | 600.0 | 602.9 | 608.7 | 620.4 |
| UD-Q4_K_M | 586.37 GiB | 588.0 | 588.7 | 590.1 | 593.0 | 598.9 | 610.5 |
| Q4_K_M | 585.99 GiB | 587.6 | 588.3 | 589.8 | 592.7 | 598.5 | 610.1 |
| Q4_K_S | 557.32 GiB | 558.9 | 559.6 | 561.1 | 564.0 | 569.8 | 581.4 |
| UD-Q4_K_S | 548.12 GiB | 549.7 | 550.4 | 551.9 | 554.8 | 560.6 | 572.2 |
| Q4_0 | 539.02 GiB | 540.6 | 541.3 | 542.8 | 545.7 | 551.5 | 563.1 |
| IQ4_NL | 537.62 GiB | 539.2 | 539.9 | 541.4 | 544.3 | 550.1 | 561.7 |
| IQ4_XS | 508.09 GiB | 509.7 | 510.4 | 511.9 | 514.8 | 520.6 | 532.2 |
| UD-IQ4_NL | 466.84 GiB | 468.4 | 469.2 | 470.6 | 473.5 | 479.3 | 490.9 |
| UD-IQ4_XS | 457.00 GiB | 458.6 | 459.3 | 460.8 | 463.7 | 469.5 | 481.1 |
| IQ3_M | 454.54 GiB | 456.1 | 456.9 | 458.3 | 461.2 | 467.0 | 478.6 |
| Q3_K_L | 452.86 GiB | 454.4 | 455.2 | 456.6 | 459.5 | 465.3 | 477.0 |
| Q3_K_M | 434.60 GiB | 436.2 | 436.9 | 438.4 | 441.3 | 447.1 | 458.7 |
| IQ3_XS | 434.58 GiB | 436.2 | 436.9 | 438.3 | 441.3 | 447.1 | 458.7 |
| UD-Q3_K_M | 428.10 GiB | 429.7 | 430.4 | 431.9 | 434.8 | 440.6 | 452.2 |
| Q3_K_S | 414.26 GiB | 415.8 | 416.6 | 418.0 | 420.9 | 426.7 | 438.4 |
| IQ3_XXS | 397.96 GiB | 399.5 | 400.3 | 401.7 | 404.6 | 410.4 | 422.1 |
| UD-IQ3_XXS | 384.32 GiB | 385.9 | 386.6 | 388.1 | 391.0 | 396.8 | 408.4 |
| UD-IQ3_S | 351.94 GiB | 353.5 | 354.3 | 355.7 | 358.6 | 364.4 | 376.0 |
| IQ3_S | 350.82 GiB | 352.4 | 353.1 | 354.6 | 357.5 | 363.3 | 374.9 |
| Q2_K_L | 334.58 GiB | 336.2 | 336.9 | 338.3 | 341.3 | 347.1 | 358.7 |
| Q2_K | 333.73 GiB | 335.3 | 336.0 | 337.5 | 340.4 | 346.2 | 357.8 |
| IQ2_M | 321.02 GiB | 322.6 | 323.3 | 324.8 | 327.7 | 333.5 | 345.1 |
| IQ2_S | 297.46 GiB | 299.0 | 299.8 | 301.2 | 304.1 | 309.9 | 321.6 |
| UD-IQ2_M | 295.48 GiB | 297.1 | 297.8 | 299.2 | 302.2 | 308.0 | 319.6 |
| UD-IQ2_XXS | 295.37 GiB | 297.0 | 297.7 | 299.1 | 302.0 | 307.9 | 319.5 |
| IQ2_XS | 285.38 GiB | 287.0 | 287.7 | 289.1 | 292.1 | 297.9 | 309.5 |
| UD-IQ1_M | 283.21 GiB | 284.8 | 285.5 | 287.0 | 289.9 | 295.7 | 307.3 |
| IQ2_XXS | 256.27 GiB | 257.9 | 258.6 | 260.0 | 262.9 | 268.8 | 280.4 |
| IQ1_M | 220.44 GiB | 222.0 | 222.8 | 224.2 | 227.1 | 232.9 | 244.5 |
| IQ1_S | 197.49 GiB | 199.1 | 199.8 | 201.3 | 204.2 | 210.0 | 221.6 |
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.