Can I run Kimi-K2-Instruct on a GeForce RTX 5060 Ti?
Not at these settings. No indexed quantization of Kimi-K2-Instruct fits GeForce RTX 5060 Ti at any context we compute, with q4_0 KV. The smallest shipped quantization is 226.87 GiB in weights alone, against 14.88 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 | 1913.1 | 1913.2 | 1913.3 | 1913.6 | 1914.2 | 1915.4 |
| Q8_0 | 1016.12 GiB | 1017.1 | 1017.2 | 1017.3 | 1017.6 | 1018.2 | 1019.4 |
| Q6_K | 784.76 GiB | 785.7 | 785.8 | 785.9 | 786.2 | 786.9 | 788.1 |
| Q5_K_M | 678.33 GiB | 679.3 | 679.4 | 679.5 | 679.8 | 680.4 | 681.6 |
| Q5_K_S | 658.04 GiB | 659.0 | 659.1 | 659.2 | 659.5 | 660.1 | 661.3 |
| Q4_1 | 598.40 GiB | 599.4 | 599.4 | 599.6 | 599.9 | 600.5 | 601.7 |
| Q4_K_M | 578.15 GiB | 579.1 | 579.2 | 579.3 | 579.6 | 580.2 | 581.4 |
| Q4_K_S | 542.73 GiB | 543.7 | 543.8 | 543.9 | 544.2 | 544.8 | 546.0 |
| Q4_0 | 540.74 GiB | 541.7 | 541.8 | 541.9 | 542.2 | 542.8 | 544.0 |
| IQ4_NL | 538.76 GiB | 539.7 | 539.8 | 539.9 | 540.2 | 540.8 | 542.1 |
| IQ4_XS | 508.98 GiB | 509.9 | 510.0 | 510.2 | 510.5 | 511.1 | 512.3 |
| Q3_K_M | 455.77 GiB | 456.7 | 456.8 | 457.0 | 457.3 | 457.9 | 459.1 |
| Q3_K_S | 412.03 GiB | 413.0 | 413.1 | 413.2 | 413.5 | 414.1 | 415.3 |
| UD-IQ3_XXS | 388.01 GiB | 389.0 | 389.0 | 389.2 | 389.5 | 390.1 | 391.3 |
| Q2_K_L | 347.81 GiB | 348.8 | 348.8 | 349.0 | 349.3 | 349.9 | 351.1 |
| Q2_K | 347.55 GiB | 348.5 | 348.6 | 348.7 | 349.0 | 349.6 | 350.8 |
| UD-IQ2_M | 323.27 GiB | 324.2 | 324.3 | 324.4 | 324.7 | 325.4 | 326.6 |
| UD-IQ2_XXS | 306.20 GiB | 307.2 | 307.2 | 307.4 | 307.7 | 308.3 | 309.5 |
| UD-IQ1_M | 283.34 GiB | 284.3 | 284.4 | 284.5 | 284.8 | 285.4 | 286.6 |
| UD-IQ1_S | 260.88 GiB | 261.8 | 261.9 | 262.1 | 262.4 | 263.0 | 264.2 |
| UD-TQ1_0 | 226.87 GiB | 227.8 | 227.9 | 228.1 | 228.4 | 229.0 | 230.2 |
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.