Can I run Kimi-K2.6 on a Radeon RX 6500 XT?
Not at these settings. No indexed quantization of Kimi-K2.6 fits Radeon RX 6500 XT at any context we compute, with q8_0 KV. The smallest shipped quantization is 193.16 GiB in weights alone, against 3.72 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 | 1913.3 | 1913.4 | 1913.7 | 1914.3 | 1915.4 | 1917.7 |
| Q4_0 | 543.62 GiB | 544.7 | 544.9 | 545.2 | 545.7 | 546.9 | 549.2 |
| IQ3_M | 454.53 GiB | 455.7 | 455.8 | 456.1 | 456.7 | 457.8 | 460.1 |
| Q3_K_L | 454.05 GiB | 455.2 | 455.3 | 455.6 | 456.2 | 457.3 | 459.6 |
| Q3_K_M | 434.81 GiB | 435.9 | 436.1 | 436.4 | 436.9 | 438.1 | 440.3 |
| IQ3_XS | 434.41 GiB | 435.5 | 435.7 | 436.0 | 436.5 | 437.7 | 439.9 |
| Q3_K_S | 413.86 GiB | 415.0 | 415.1 | 415.4 | 416.0 | 417.1 | 419.4 |
| IQ3_XXS | 397.34 GiB | 398.5 | 398.6 | 398.9 | 399.5 | 400.6 | 402.9 |
| Q2_K_L | 334.66 GiB | 335.8 | 335.9 | 336.2 | 336.8 | 337.9 | 340.2 |
| Q2_K | 333.59 GiB | 334.7 | 334.9 | 335.1 | 335.7 | 336.8 | 339.1 |
| IQ2_M | 318.27 GiB | 319.4 | 319.5 | 319.8 | 320.4 | 321.5 | 323.8 |
| IQ2_S | 287.58 GiB | 288.7 | 288.8 | 289.1 | 289.7 | 290.8 | 293.1 |
| IQ2_XS | 282.35 GiB | 283.5 | 283.6 | 283.9 | 284.5 | 285.6 | 287.9 |
| IQ2_XXS | 252.81 GiB | 253.9 | 254.1 | 254.4 | 254.9 | 256.1 | 258.3 |
| IQ1_M | 216.46 GiB | 217.6 | 217.7 | 218.0 | 218.6 | 219.7 | 222.0 |
| IQ1_S | 193.16 GiB | 194.3 | 194.4 | 194.7 | 195.3 | 196.4 | 198.7 |
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.