Can I run Kimi-K2-Thinking on a GeForce RTX 3050?
Not at these settings. No indexed quantization of Kimi-K2-Thinking fits GeForce RTX 3050 at any context we compute, with q8_0 KV. The smallest shipped quantization is 229.90 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 | 1912.15 GiB | 1913.2 | 1913.3 | 1913.6 | 1914.2 | 1915.3 | 1917.6 |
| Q8_0 | 1016.12 GiB | 1017.1 | 1017.3 | 1017.6 | 1018.1 | 1019.3 | 1021.6 |
| Q6_K | 785.02 GiB | 786.0 | 786.2 | 786.5 | 787.0 | 788.2 | 790.5 |
| Q5_K_M | 678.69 GiB | 679.7 | 679.9 | 680.1 | 680.7 | 681.8 | 684.1 |
| Q5_K_S | 658.45 GiB | 659.5 | 659.6 | 659.9 | 660.5 | 661.6 | 663.9 |
| Q4_1 | 598.76 GiB | 599.8 | 599.9 | 600.2 | 600.8 | 601.9 | 604.2 |
| Q4_K_M | 578.58 GiB | 579.6 | 579.7 | 580.0 | 580.6 | 581.7 | 584.0 |
| Q4_K_S | 543.25 GiB | 544.3 | 544.4 | 544.7 | 545.3 | 546.4 | 548.7 |
| Q4_0 | 541.27 GiB | 542.3 | 542.4 | 542.7 | 543.3 | 544.4 | 546.7 |
| IQ4_NL | 539.30 GiB | 540.3 | 540.5 | 540.7 | 541.3 | 542.5 | 544.7 |
| IQ4_XS | 509.60 GiB | 510.6 | 510.8 | 511.0 | 511.6 | 512.8 | 515.0 |
| Q3_K_M | 456.34 GiB | 457.4 | 457.5 | 457.8 | 458.4 | 459.5 | 461.8 |
| Q3_K_S | 412.72 GiB | 413.7 | 413.9 | 414.2 | 414.7 | 415.9 | 418.2 |
| UD-IQ3_XXS | 392.74 GiB | 393.8 | 393.9 | 394.2 | 394.8 | 395.9 | 398.2 |
| IQ3_XXS | 367.09 GiB | 368.1 | 368.3 | 368.5 | 369.1 | 370.2 | 372.5 |
| Q2_K_L | 348.68 GiB | 349.7 | 349.8 | 350.1 | 350.7 | 351.8 | 354.1 |
| Q2_K | 348.43 GiB | 349.4 | 349.6 | 349.9 | 350.4 | 351.6 | 353.9 |
| UD-IQ2_M | 329.27 GiB | 330.3 | 330.4 | 330.7 | 331.3 | 332.4 | 334.7 |
| UD-IQ2_XXS | 312.21 GiB | 313.2 | 313.4 | 313.7 | 314.2 | 315.4 | 317.6 |
| UD-IQ1_M | 288.02 GiB | 289.0 | 289.2 | 289.5 | 290.0 | 291.2 | 293.5 |
| IQ2_S | 280.26 GiB | 281.3 | 281.4 | 281.7 | 282.3 | 283.4 | 285.7 |
| IQ2_XS | 278.07 GiB | 279.1 | 279.2 | 279.5 | 280.1 | 281.2 | 283.5 |
| UD-IQ1_S | 265.74 GiB | 266.8 | 266.9 | 267.2 | 267.8 | 268.9 | 271.2 |
| UD-TQ1_0 | 229.90 GiB | 230.9 | 231.1 | 231.3 | 231.9 | 233.1 | 235.3 |
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, this model uses latent attention and allocates no V cache at all, so any formula reading num_key_value_heads overstates its cache by more than an order of magnitude.