Can I run Step-3.5-Flash-REAP-121B-A11B on a Apple M3 Pro?
Not at these settings. No indexed quantization of Step-3.5-Flash-REAP-121B-A11B fits Apple M3 Pro at any context we compute, with q8_0 KV. The smallest shipped quantization is 22.74 GiB in weights alone, against 12.56 GiB usable. CPU offload can still run it, slowly.
Every quantization at every context
| Quant | Weights● | 4K◐ | 8K◐ | 16K◐ | 32K◐ | 64K◐ | 128K◐ |
|---|---|---|---|---|---|---|---|
| I1-Q6_K | 92.51 GiB | 94.4 | 95.2 | 96.8 | 100.0 | 106.4 | 119.1 |
| I1-Q5_K_M | 79.79 GiB | 81.7 | 82.5 | 84.1 | 87.3 | 93.7 | 106.4 |
| I1-Q5_K_S | 77.62 GiB | 79.5 | 80.3 | 81.9 | 85.1 | 91.5 | 104.2 |
| I1-Q4_1 | 70.61 GiB | 72.5 | 73.3 | 74.9 | 78.1 | 84.5 | 97.2 |
| I1-Q4_K_M | 67.82 GiB | 69.7 | 70.5 | 72.1 | 75.3 | 81.7 | 94.4 |
| I1-Q4_K_S | 63.83 GiB | 65.8 | 66.5 | 68.1 | 71.3 | 77.7 | 90.5 |
| I1-Q4_0 | 63.71 GiB | 65.6 | 66.4 | 68.0 | 71.2 | 77.6 | 90.3 |
| I1-IQ4_XS | 60.12 GiB | 62.0 | 62.8 | 64.4 | 67.6 | 74.0 | 86.7 |
| I1-Q3_K_L | 58.48 GiB | 60.4 | 61.2 | 62.8 | 66.0 | 72.4 | 85.1 |
| I1-Q3_K_M | 53.75 GiB | 55.7 | 56.5 | 58.1 | 61.3 | 67.6 | 80.4 |
| I1-IQ3_M | 49.23 GiB | 51.1 | 51.9 | 53.5 | 56.7 | 63.1 | 75.9 |
| I1-IQ3_S | 48.73 GiB | 50.7 | 51.4 | 53.0 | 56.2 | 62.6 | 75.4 |
| I1-Q3_K_S | 48.71 GiB | 50.6 | 51.4 | 53.0 | 56.2 | 62.6 | 75.3 |
| I1-IQ3_XS | 45.92 GiB | 47.8 | 48.6 | 50.2 | 53.4 | 59.8 | 72.5 |
| I1-IQ3_XXS | 43.40 GiB | 45.3 | 46.1 | 47.7 | 50.9 | 57.3 | 70.0 |
| I1-Q2_K | 41.19 GiB | 43.1 | 43.9 | 45.5 | 48.7 | 55.1 | 67.8 |
| I1-Q2_K_S | 37.79 GiB | 39.7 | 40.5 | 42.1 | 45.3 | 51.7 | 64.4 |
| I1-IQ2_M | 36.79 GiB | 38.7 | 39.5 | 41.1 | 44.3 | 50.7 | 63.4 |
| I1-IQ2_S | 33.40 GiB | 35.3 | 36.1 | 37.7 | 40.9 | 47.3 | 60.0 |
| I1-IQ2_XS | 32.98 GiB | 34.9 | 35.7 | 37.3 | 40.5 | 46.9 | 59.6 |
| I1-IQ2_XXS | 29.52 GiB | 31.4 | 32.2 | 33.8 | 37.0 | 43.4 | 56.1 |
| I1-IQ1_M | 25.28 GiB | 27.2 | 28.0 | 29.6 | 32.8 | 39.2 | 51.9 |
| I1-IQ1_S | 22.74 GiB | 24.7 | 25.5 | 27.1 | 30.2 | 36.6 | 49.4 |
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, most of this model's layers cache only a 512-token window rather than the full context.