Can I run Kimi-K2.7-Code on a Apple M5 Pro?
Not at these settings. No indexed quantization of Kimi-K2.7-Code fits Apple M5 Pro at any context we compute, with q4_0 KV. The smallest shipped quantization is 283.04 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◐ |
|---|---|---|---|---|---|---|---|
| UD-IQ4_XS | 461.08 GiB | 461.8 | 461.9 | 462.0 | 462.3 | 462.9 | 464.1 |
| UD-Q3_K_M | 431.78 GiB | 432.5 | 432.6 | 432.7 | 433.0 | 433.6 | 434.8 |
| UD-IQ3_S | 390.04 GiB | 390.7 | 390.8 | 391.0 | 391.3 | 391.9 | 393.1 |
| UD-IQ3_XXS | 351.00 GiB | 351.7 | 351.8 | 351.9 | 352.2 | 352.8 | 354.0 |
| UD-IQ2_M | 296.14 GiB | 296.8 | 296.9 | 297.1 | 297.4 | 298.0 | 299.2 |
| UD-IQ2_XXS | 296.00 GiB | 296.7 | 296.8 | 296.9 | 297.2 | 297.8 | 299.0 |
| UD-IQ1_M | 283.04 GiB | 283.7 | 283.8 | 284.0 | 284.3 | 284.9 | 286.1 |
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.