Can I run Step-3.7-Flash on a Apple M5 Pro?
Not at these settings. No indexed quantization of Step-3.7-Flash fits Apple M5 Pro at any context we compute, with q4_0 KV. The smallest shipped quantization is 40.03 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 | 366.96 GiB | 368.2 | 368.7 | 369.5 | 371.2 | 374.6 | 381.3 |
| Q8_0 | 197.44 GiB | 198.7 | 199.2 | 200.0 | 201.7 | 205.1 | 211.8 |
| Q6_K_L | 160.20 GiB | 161.5 | 161.9 | 162.8 | 164.4 | 167.8 | 174.6 |
| Q6_K | 159.96 GiB | 161.3 | 161.7 | 162.5 | 164.2 | 167.6 | 174.3 |
| UD-Q6_K | 152.12 GiB | 153.4 | 153.8 | 154.7 | 156.4 | 159.7 | 166.5 |
| UD-Q5_K_M | 136.43 GiB | 137.7 | 138.1 | 139.0 | 140.7 | 144.1 | 150.8 |
| Q5_K_L | 132.58 GiB | 133.9 | 134.3 | 135.1 | 136.8 | 140.2 | 147.0 |
| Q5_K_M | 132.28 GiB | 133.6 | 134.0 | 134.8 | 136.5 | 139.9 | 146.6 |
| UD-Q5_K_S | 128.59 GiB | 129.9 | 130.3 | 131.1 | 132.8 | 136.2 | 143.0 |
| Q5_K_S | 128.01 GiB | 129.3 | 129.7 | 130.6 | 132.3 | 135.6 | 142.4 |
| Q4_1 | 116.67 GiB | 118.0 | 118.4 | 119.2 | 120.9 | 124.3 | 131.0 |
| UD-Q4_K_M | 113.71 GiB | 115.0 | 115.4 | 116.3 | 117.9 | 121.3 | 128.1 |
| Q4_K_L | 113.63 GiB | 114.9 | 115.3 | 116.2 | 117.9 | 121.3 | 128.0 |
| Q4_K_M | 113.27 GiB | 114.6 | 115.0 | 115.8 | 117.5 | 120.9 | 127.6 |
| Q4_K_S | 109.05 GiB | 110.3 | 110.8 | 111.6 | 113.3 | 116.7 | 123.4 |
| UD-Q4_K_S | 106.32 GiB | 107.6 | 108.0 | 108.9 | 110.6 | 113.9 | 120.7 |
| Q4_0 | 105.79 GiB | 107.1 | 107.5 | 108.3 | 110.0 | 113.4 | 120.2 |
| IQ4_NL | 105.54 GiB | 106.8 | 107.3 | 108.1 | 109.8 | 113.2 | 119.9 |
| IQ4_XS | 99.94 GiB | 101.2 | 101.6 | 102.5 | 104.2 | 107.6 | 114.3 |
| UD-IQ4_NL | 90.63 GiB | 91.9 | 92.3 | 93.2 | 94.9 | 98.3 | 105.0 |
| IQ3_M | 89.38 GiB | 90.7 | 91.1 | 91.9 | 93.6 | 97.0 | 103.8 |
| Q3_K_L | 88.94 GiB | 90.2 | 90.7 | 91.5 | 93.2 | 96.6 | 103.3 |
| UD-IQ4_XS | 88.79 GiB | 90.1 | 90.5 | 91.3 | 93.0 | 96.4 | 103.2 |
| Q3_K_M | 87.36 GiB | 88.6 | 89.1 | 89.9 | 91.6 | 95.0 | 101.7 |
| IQ3_XS | 85.44 GiB | 86.7 | 87.1 | 88.0 | 89.7 | 93.1 | 99.8 |
| UD-Q3_K_M | 83.13 GiB | 84.4 | 84.8 | 85.7 | 87.4 | 90.8 | 97.5 |
| Q3_K_S | 81.55 GiB | 82.8 | 83.3 | 84.1 | 85.8 | 89.2 | 95.9 |
| IQ3_XXS | 78.38 GiB | 79.7 | 80.1 | 80.9 | 82.6 | 86.0 | 92.8 |
| UD-IQ3_S | 74.54 GiB | 75.8 | 76.2 | 77.1 | 78.8 | 82.2 | 88.9 |
| UD-IQ3_XXS | 68.54 GiB | 69.8 | 70.2 | 71.1 | 72.8 | 76.2 | 82.9 |
| Q2_K_L | 66.82 GiB | 68.1 | 68.5 | 69.4 | 71.1 | 74.4 | 81.2 |
| Q2_K | 66.34 GiB | 67.6 | 68.0 | 68.9 | 70.6 | 74.0 | 80.7 |
| IQ2_M | 63.68 GiB | 65.0 | 65.4 | 66.2 | 67.9 | 71.3 | 78.1 |
| IQ2_S | 57.93 GiB | 59.2 | 59.6 | 60.5 | 62.2 | 65.5 | 72.3 |
| UD-IQ2_M | 57.58 GiB | 58.9 | 59.3 | 60.1 | 61.8 | 65.2 | 71.9 |
| UD-IQ2_XXS | 57.47 GiB | 58.8 | 59.2 | 60.0 | 61.7 | 65.1 | 71.8 |
| IQ2_XS | 56.76 GiB | 58.0 | 58.5 | 59.3 | 61.0 | 64.4 | 71.1 |
| UD-IQ1_M | 52.85 GiB | 54.1 | 54.6 | 55.4 | 57.1 | 60.5 | 67.2 |
| IQ2_XXS | 51.22 GiB | 52.5 | 52.9 | 53.8 | 55.5 | 58.8 | 65.6 |
| IQ1_M | 44.41 GiB | 45.7 | 46.1 | 47.0 | 48.6 | 52.0 | 58.8 |
| IQ1_S | 40.03 GiB | 41.3 | 41.7 | 42.6 | 44.3 | 47.7 | 54.4 |
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.