Can I run Kimi-K2-Instruct on a Radeon RX 6500 XT?
Not at these settings. No indexed quantization of Kimi-K2-Instruct fits Radeon RX 6500 XT at any context we compute, with q4_0 KV. The smallest shipped quantization is 226.87 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 | 1913.2 | 1913.3 | 1913.4 | 1913.7 | 1914.3 | 1915.5 |
| Q8_0 | 1016.12 GiB | 1017.2 | 1017.3 | 1017.4 | 1017.7 | 1018.3 | 1019.5 |
| Q6_K | 784.76 GiB | 785.8 | 785.9 | 786.0 | 786.3 | 787.0 | 788.2 |
| Q5_K_M | 678.33 GiB | 679.4 | 679.5 | 679.6 | 679.9 | 680.5 | 681.7 |
| Q5_K_S | 658.04 GiB | 659.1 | 659.2 | 659.3 | 659.6 | 660.2 | 661.4 |
| Q4_1 | 598.40 GiB | 599.5 | 599.5 | 599.7 | 600.0 | 600.6 | 601.8 |
| Q4_K_M | 578.15 GiB | 579.2 | 579.3 | 579.4 | 579.7 | 580.3 | 581.5 |
| Q4_K_S | 542.73 GiB | 543.8 | 543.9 | 544.0 | 544.3 | 544.9 | 546.1 |
| Q4_0 | 540.74 GiB | 541.8 | 541.9 | 542.0 | 542.3 | 542.9 | 544.1 |
| IQ4_NL | 538.76 GiB | 539.8 | 539.9 | 540.0 | 540.3 | 540.9 | 542.2 |
| IQ4_XS | 508.98 GiB | 510.0 | 510.1 | 510.3 | 510.6 | 511.2 | 512.4 |
| Q3_K_M | 455.77 GiB | 456.8 | 456.9 | 457.1 | 457.4 | 458.0 | 459.2 |
| Q3_K_S | 412.03 GiB | 413.1 | 413.2 | 413.3 | 413.6 | 414.2 | 415.4 |
| UD-IQ3_XXS | 388.01 GiB | 389.1 | 389.1 | 389.3 | 389.6 | 390.2 | 391.4 |
| Q2_K_L | 347.81 GiB | 348.9 | 348.9 | 349.1 | 349.4 | 350.0 | 351.2 |
| Q2_K | 347.55 GiB | 348.6 | 348.7 | 348.8 | 349.1 | 349.7 | 350.9 |
| UD-IQ2_M | 323.27 GiB | 324.3 | 324.4 | 324.5 | 324.8 | 325.5 | 326.7 |
| UD-IQ2_XXS | 306.20 GiB | 307.3 | 307.3 | 307.5 | 307.8 | 308.4 | 309.6 |
| UD-IQ1_M | 283.34 GiB | 284.4 | 284.5 | 284.6 | 284.9 | 285.5 | 286.7 |
| UD-IQ1_S | 260.88 GiB | 261.9 | 262.0 | 262.2 | 262.5 | 263.1 | 264.3 |
| UD-TQ1_0 | 226.87 GiB | 227.9 | 228.0 | 228.2 | 228.5 | 229.1 | 230.3 |
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.