Can I run Kimi-K2-Thinking on a Radeon RX 6500 XT?
Not at these settings. No indexed quantization of Kimi-K2-Thinking fits Radeon RX 6500 XT at any context we compute, with q8_0 KV. The smallest shipped quantization is 229.90 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 |
| Q8_0 | 1016.12 GiB | 1017.2 | 1017.4 | 1017.7 | 1018.2 | 1019.4 | 1021.7 |
| Q6_K | 785.02 GiB | 786.1 | 786.3 | 786.6 | 787.1 | 788.3 | 790.6 |
| Q5_K_M | 678.69 GiB | 679.8 | 680.0 | 680.2 | 680.8 | 681.9 | 684.2 |
| Q5_K_S | 658.45 GiB | 659.6 | 659.7 | 660.0 | 660.6 | 661.7 | 664.0 |
| Q4_1 | 598.76 GiB | 599.9 | 600.0 | 600.3 | 600.9 | 602.0 | 604.3 |
| Q4_K_M | 578.58 GiB | 579.7 | 579.8 | 580.1 | 580.7 | 581.8 | 584.1 |
| Q4_K_S | 543.25 GiB | 544.4 | 544.5 | 544.8 | 545.4 | 546.5 | 548.8 |
| Q4_0 | 541.27 GiB | 542.4 | 542.5 | 542.8 | 543.4 | 544.5 | 546.8 |
| IQ4_NL | 539.30 GiB | 540.4 | 540.6 | 540.8 | 541.4 | 542.6 | 544.8 |
| IQ4_XS | 509.60 GiB | 510.7 | 510.9 | 511.1 | 511.7 | 512.9 | 515.1 |
| Q3_K_M | 456.34 GiB | 457.5 | 457.6 | 457.9 | 458.5 | 459.6 | 461.9 |
| Q3_K_S | 412.72 GiB | 413.8 | 414.0 | 414.3 | 414.8 | 416.0 | 418.3 |
| UD-IQ3_XXS | 392.74 GiB | 393.9 | 394.0 | 394.3 | 394.9 | 396.0 | 398.3 |
| IQ3_XXS | 367.09 GiB | 368.2 | 368.4 | 368.6 | 369.2 | 370.3 | 372.6 |
| Q2_K_L | 348.68 GiB | 349.8 | 349.9 | 350.2 | 350.8 | 351.9 | 354.2 |
| Q2_K | 348.43 GiB | 349.5 | 349.7 | 350.0 | 350.5 | 351.7 | 354.0 |
| UD-IQ2_M | 329.27 GiB | 330.4 | 330.5 | 330.8 | 331.4 | 332.5 | 334.8 |
| UD-IQ2_XXS | 312.21 GiB | 313.3 | 313.5 | 313.8 | 314.3 | 315.5 | 317.7 |
| UD-IQ1_M | 288.02 GiB | 289.1 | 289.3 | 289.6 | 290.1 | 291.3 | 293.6 |
| IQ2_S | 280.26 GiB | 281.4 | 281.5 | 281.8 | 282.4 | 283.5 | 285.8 |
| IQ2_XS | 278.07 GiB | 279.2 | 279.3 | 279.6 | 280.2 | 281.3 | 283.6 |
| UD-IQ1_S | 265.74 GiB | 266.9 | 267.0 | 267.3 | 267.9 | 269.0 | 271.3 |
| UD-TQ1_0 | 229.90 GiB | 231.0 | 231.2 | 231.4 | 232.0 | 233.2 | 235.4 |
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.