Can I run Kimi-K2.6 on a GeForce RTX 3050?
Not at these settings. No indexed quantization of Kimi-K2.6 fits GeForce RTX 3050 at any context we compute, with q8_0 KV. The smallest shipped quantization is 193.16 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 |
| Q4_0 | 543.62 GiB | 544.6 | 544.8 | 545.1 | 545.6 | 546.8 | 549.1 |
| IQ3_M | 454.53 GiB | 455.6 | 455.7 | 456.0 | 456.6 | 457.7 | 460.0 |
| Q3_K_L | 454.05 GiB | 455.1 | 455.2 | 455.5 | 456.1 | 457.2 | 459.5 |
| Q3_K_M | 434.81 GiB | 435.8 | 436.0 | 436.3 | 436.8 | 438.0 | 440.2 |
| IQ3_XS | 434.41 GiB | 435.4 | 435.6 | 435.9 | 436.4 | 437.6 | 439.8 |
| Q3_K_S | 413.86 GiB | 414.9 | 415.0 | 415.3 | 415.9 | 417.0 | 419.3 |
| IQ3_XXS | 397.34 GiB | 398.4 | 398.5 | 398.8 | 399.4 | 400.5 | 402.8 |
| Q2_K_L | 334.66 GiB | 335.7 | 335.8 | 336.1 | 336.7 | 337.8 | 340.1 |
| Q2_K | 333.59 GiB | 334.6 | 334.8 | 335.0 | 335.6 | 336.7 | 339.0 |
| IQ2_M | 318.27 GiB | 319.3 | 319.4 | 319.7 | 320.3 | 321.4 | 323.7 |
| IQ2_S | 287.58 GiB | 288.6 | 288.7 | 289.0 | 289.6 | 290.7 | 293.0 |
| IQ2_XS | 282.35 GiB | 283.4 | 283.5 | 283.8 | 284.4 | 285.5 | 287.8 |
| IQ2_XXS | 252.81 GiB | 253.8 | 254.0 | 254.3 | 254.8 | 256.0 | 258.2 |
| IQ1_M | 216.46 GiB | 217.5 | 217.6 | 217.9 | 218.5 | 219.6 | 221.9 |
| IQ1_S | 193.16 GiB | 194.2 | 194.3 | 194.6 | 195.2 | 196.3 | 198.6 |
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, 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.