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 q4_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 | 93.8 | 94.2 | 95.1 | 96.8 | 100.1 | 106.9 |
| I1-Q5_K_M | 79.79 GiB | 81.1 | 81.5 | 82.3 | 84.0 | 87.4 | 94.2 |
| I1-Q5_K_S | 77.62 GiB | 78.9 | 79.3 | 80.2 | 81.9 | 85.2 | 92.0 |
| I1-Q4_1 | 70.61 GiB | 71.9 | 72.3 | 73.2 | 74.9 | 78.2 | 85.0 |
| I1-Q4_K_M | 67.82 GiB | 69.1 | 69.5 | 70.4 | 72.1 | 75.4 | 82.2 |
| I1-Q4_K_S | 63.83 GiB | 65.1 | 65.5 | 66.4 | 68.1 | 71.4 | 78.2 |
| I1-Q4_0 | 63.71 GiB | 65.0 | 65.4 | 66.3 | 68.0 | 71.3 | 78.1 |
| I1-IQ4_XS | 60.12 GiB | 61.4 | 61.8 | 62.7 | 64.4 | 67.7 | 74.5 |
| I1-Q3_K_L | 58.48 GiB | 59.8 | 60.2 | 61.0 | 62.7 | 66.1 | 72.8 |
| I1-Q3_K_M | 53.75 GiB | 55.0 | 55.5 | 56.3 | 58.0 | 61.4 | 68.1 |
| I1-IQ3_M | 49.23 GiB | 50.5 | 50.9 | 51.8 | 53.5 | 56.8 | 63.6 |
| I1-IQ3_S | 48.73 GiB | 50.0 | 50.4 | 51.3 | 53.0 | 56.3 | 63.1 |
| I1-Q3_K_S | 48.71 GiB | 50.0 | 50.4 | 51.3 | 52.9 | 56.3 | 63.1 |
| I1-IQ3_XS | 45.92 GiB | 47.2 | 47.6 | 48.5 | 50.2 | 53.5 | 60.3 |
| I1-IQ3_XXS | 43.40 GiB | 44.7 | 45.1 | 46.0 | 47.6 | 51.0 | 57.8 |
| I1-Q2_K | 41.19 GiB | 42.5 | 42.9 | 43.7 | 45.4 | 48.8 | 55.6 |
| I1-Q2_K_S | 37.79 GiB | 39.1 | 39.5 | 40.3 | 42.0 | 45.4 | 52.2 |
| I1-IQ2_M | 36.79 GiB | 38.1 | 38.5 | 39.3 | 41.0 | 44.4 | 51.2 |
| I1-IQ2_S | 33.40 GiB | 34.7 | 35.1 | 36.0 | 37.6 | 41.0 | 47.8 |
| I1-IQ2_XS | 32.98 GiB | 34.3 | 34.7 | 35.5 | 37.2 | 40.6 | 47.3 |
| I1-IQ2_XXS | 29.52 GiB | 30.8 | 31.2 | 32.1 | 33.8 | 37.1 | 43.9 |
| I1-IQ1_M | 25.28 GiB | 26.6 | 27.0 | 27.8 | 29.5 | 32.9 | 39.7 |
| I1-IQ1_S | 22.74 GiB | 24.0 | 24.4 | 25.3 | 27.0 | 30.4 | 37.1 |
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, most of this model's layers cache only a 512-token window rather than the full context.