Can I run Kimi-K2-Thinking on a Apple M3 Ultra?
Not at these settings. No indexed quantization of Kimi-K2-Thinking fits Apple M3 Ultra at any context we compute, with q4_0 KV. The smallest shipped quantization is 229.90 GiB in weights alone, against 66.96 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 | 1912.9 | 1913.1 | 1913.4 | 1914.0 | 1915.2 |
| Q8_0 | 1016.12 GiB | 1016.8 | 1016.9 | 1017.1 | 1017.4 | 1018.0 | 1019.2 |
| Q6_K | 785.02 GiB | 785.7 | 785.8 | 786.0 | 786.3 | 786.9 | 788.1 |
| Q5_K_M | 678.69 GiB | 679.4 | 679.5 | 679.6 | 679.9 | 680.5 | 681.7 |
| Q5_K_S | 658.45 GiB | 659.2 | 659.2 | 659.4 | 659.7 | 660.3 | 661.5 |
| Q4_1 | 598.76 GiB | 599.5 | 599.5 | 599.7 | 600.0 | 600.6 | 601.8 |
| Q4_K_M | 578.58 GiB | 579.3 | 579.4 | 579.5 | 579.8 | 580.4 | 581.6 |
| Q4_K_S | 543.25 GiB | 544.0 | 544.0 | 544.2 | 544.5 | 545.1 | 546.3 |
| Q4_0 | 541.27 GiB | 542.0 | 542.1 | 542.2 | 542.5 | 543.1 | 544.3 |
| IQ4_NL | 539.30 GiB | 540.0 | 540.1 | 540.2 | 540.5 | 541.1 | 542.3 |
| IQ4_XS | 509.60 GiB | 510.3 | 510.4 | 510.5 | 510.8 | 511.4 | 512.6 |
| Q3_K_M | 456.34 GiB | 457.0 | 457.1 | 457.3 | 457.6 | 458.2 | 459.4 |
| Q3_K_S | 412.72 GiB | 413.4 | 413.5 | 413.7 | 414.0 | 414.6 | 415.8 |
| UD-IQ3_XXS | 392.74 GiB | 393.4 | 393.5 | 393.7 | 394.0 | 394.6 | 395.8 |
| IQ3_XXS | 367.09 GiB | 367.8 | 367.9 | 368.0 | 368.3 | 368.9 | 370.1 |
| Q2_K_L | 348.68 GiB | 349.4 | 349.5 | 349.6 | 349.9 | 350.5 | 351.7 |
| Q2_K | 348.43 GiB | 349.1 | 349.2 | 349.4 | 349.7 | 350.3 | 351.5 |
| UD-IQ2_M | 329.27 GiB | 330.0 | 330.1 | 330.2 | 330.5 | 331.1 | 332.3 |
| UD-IQ2_XXS | 312.21 GiB | 312.9 | 313.0 | 313.1 | 313.4 | 314.0 | 315.3 |
| UD-IQ1_M | 288.02 GiB | 288.7 | 288.8 | 289.0 | 289.3 | 289.9 | 291.1 |
| IQ2_S | 280.26 GiB | 281.0 | 281.0 | 281.2 | 281.5 | 282.1 | 283.3 |
| IQ2_XS | 278.07 GiB | 278.8 | 278.8 | 279.0 | 279.3 | 279.9 | 281.1 |
| UD-IQ1_S | 265.74 GiB | 266.4 | 266.5 | 266.7 | 267.0 | 267.6 | 268.8 |
| UD-TQ1_0 | 229.90 GiB | 230.6 | 230.7 | 230.8 | 231.1 | 231.7 | 232.9 |
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.