Can I run Kimi-K2.5 on a Radeon RX 6500 XT?
Not at these settings. No indexed quantization of Kimi-K2.5 fits Radeon RX 6500 XT at any context we compute, with q8_0 KV. The smallest shipped quantization is 195.86 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 |
| 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 | 598.99 GiB | 600.1 | 600.3 | 600.5 | 601.1 | 602.2 | 604.5 |
| Q4_K_M | 578.58 GiB | 579.7 | 579.8 | 580.1 | 580.7 | 581.8 | 584.1 |
| Q4_0 | 549.26 GiB | 550.4 | 550.5 | 550.8 | 551.4 | 552.5 | 554.8 |
| Q8_0 | 543.62 GiB | 544.7 | 544.9 | 545.2 | 545.7 | 546.9 | 549.2 |
| Q4_K_S | 543.24 GiB | 544.4 | 544.5 | 544.8 | 545.4 | 546.5 | 548.8 |
| Q4_K_L | 541.31 GiB | 542.4 | 542.6 | 542.9 | 543.4 | 544.6 | 546.8 |
| IQ4_NL | 539.67 GiB | 540.8 | 540.9 | 541.2 | 541.8 | 542.9 | 545.2 |
| IQ4_XS | 510.00 GiB | 511.1 | 511.3 | 511.5 | 512.1 | 513.3 | 515.5 |
| Q3_K_M | 456.14 GiB | 457.3 | 457.4 | 457.7 | 458.3 | 459.4 | 461.7 |
| Q3_K_L | 454.05 GiB | 455.2 | 455.3 | 455.6 | 456.2 | 457.3 | 459.6 |
| IQ3_M | 434.81 GiB | 435.9 | 436.1 | 436.4 | 436.9 | 438.1 | 440.3 |
| Q3_K_S | 414.28 GiB | 415.4 | 415.5 | 415.8 | 416.4 | 417.5 | 419.8 |
| IQ3_XS | 391.23 GiB | 392.4 | 392.5 | 392.8 | 393.4 | 394.5 | 396.8 |
| UD-IQ3_XXS | 386.32 GiB | 387.4 | 387.6 | 387.9 | 388.4 | 389.6 | 391.9 |
| IQ3_S | 377.51 GiB | 378.6 | 378.8 | 379.1 | 379.6 | 380.8 | 383.0 |
| IQ3_XXS | 376.80 GiB | 377.9 | 378.1 | 378.3 | 378.9 | 380.1 | 382.3 |
| Q2_K_L | 348.36 GiB | 349.5 | 349.6 | 349.9 | 350.5 | 351.6 | 353.9 |
| Q2_K | 348.11 GiB | 349.2 | 349.4 | 349.7 | 350.2 | 351.4 | 353.6 |
| UD-IQ2_M | 321.54 GiB | 322.7 | 322.8 | 323.1 | 323.7 | 324.8 | 327.1 |
| IQ2_S | 311.72 GiB | 312.8 | 313.0 | 313.3 | 313.8 | 315.0 | 317.3 |
| UD-IQ2_XXS | 304.30 GiB | 305.4 | 305.6 | 305.8 | 306.4 | 307.6 | 309.8 |
| IQ2_M | 300.77 GiB | 301.9 | 302.0 | 302.3 | 302.9 | 304.0 | 306.3 |
| UD-IQ1_M | 279.94 GiB | 281.1 | 281.2 | 281.5 | 282.1 | 283.2 | 285.5 |
| IQ2_XS | 263.67 GiB | 264.8 | 264.9 | 265.2 | 265.8 | 266.9 | 269.2 |
| IQ2_XXS | 262.75 GiB | 263.9 | 264.0 | 264.3 | 264.9 | 266.0 | 268.3 |
| UD-IQ1_S | 256.97 GiB | 258.1 | 258.2 | 258.5 | 259.1 | 260.2 | 262.5 |
| UD-TQ1_0 | 223.09 GiB | 224.2 | 224.4 | 224.6 | 225.2 | 226.3 | 228.6 |
| IQ1_M | 204.50 GiB | 205.6 | 205.8 | 206.1 | 206.6 | 207.8 | 210.0 |
| IQ1_S | 195.86 GiB | 197.0 | 197.1 | 197.4 | 198.0 | 199.1 | 201.4 |
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.