Can I run Kimi-K2.6 on a Apple M5 Pro?
Not at these settings. No indexed quantization of Kimi-K2.6 fits Apple M5 Pro at any context we compute, with q8_0 KV. The smallest shipped quantization is 193.16 GiB in weights alone, against 33.48 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 | 1912.9 | 1913.1 | 1913.3 | 1913.9 | 1915.1 | 1917.3 |
| Q4_0 | 543.62 GiB | 544.4 | 544.5 | 544.8 | 545.4 | 546.5 | 548.8 |
| IQ3_M | 454.53 GiB | 455.3 | 455.4 | 455.7 | 456.3 | 457.4 | 459.7 |
| Q3_K_L | 454.05 GiB | 454.8 | 455.0 | 455.3 | 455.8 | 457.0 | 459.2 |
| Q3_K_M | 434.81 GiB | 435.6 | 435.7 | 436.0 | 436.6 | 437.7 | 440.0 |
| IQ3_XS | 434.41 GiB | 435.2 | 435.3 | 435.6 | 436.2 | 437.3 | 439.6 |
| Q3_K_S | 413.86 GiB | 414.6 | 414.8 | 415.1 | 415.6 | 416.8 | 419.0 |
| IQ3_XXS | 397.34 GiB | 398.1 | 398.2 | 398.5 | 399.1 | 400.2 | 402.5 |
| Q2_K_L | 334.66 GiB | 335.4 | 335.6 | 335.9 | 336.4 | 337.6 | 339.8 |
| Q2_K | 333.59 GiB | 334.4 | 334.5 | 334.8 | 335.4 | 336.5 | 338.8 |
| IQ2_M | 318.27 GiB | 319.0 | 319.2 | 319.5 | 320.0 | 321.2 | 323.5 |
| IQ2_S | 287.58 GiB | 288.4 | 288.5 | 288.8 | 289.4 | 290.5 | 292.8 |
| IQ2_XS | 282.35 GiB | 283.1 | 283.3 | 283.6 | 284.1 | 285.3 | 287.5 |
| IQ2_XXS | 252.81 GiB | 253.6 | 253.7 | 254.0 | 254.6 | 255.7 | 258.0 |
| IQ1_M | 216.46 GiB | 217.2 | 217.4 | 217.7 | 218.2 | 219.4 | 221.7 |
| IQ1_S | 193.16 GiB | 193.9 | 194.1 | 194.4 | 194.9 | 196.1 | 198.3 |
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.