Can I run Step-3.7-Flash on a Apple M5 Max?
Not at these settings. No indexed quantization of Step-3.7-Flash fits Apple M5 Max at any context we compute, with q8_0 KV. The smallest shipped quantization is 40.03 GiB in weights alone, against 25.11 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 | 368.9 | 369.7 | 371.3 | 374.5 | 380.8 | 393.6 |
| Q8_0 | 197.44 GiB | 199.4 | 200.2 | 201.8 | 204.9 | 211.3 | 224.1 |
| Q6_K_L | 160.20 GiB | 162.1 | 162.9 | 164.5 | 167.7 | 174.1 | 186.8 |
| Q6_K | 159.96 GiB | 161.9 | 162.7 | 164.3 | 167.5 | 173.8 | 186.6 |
| UD-Q6_K | 152.12 GiB | 154.0 | 154.8 | 156.4 | 159.6 | 166.0 | 178.7 |
| UD-Q5_K_M | 136.43 GiB | 138.4 | 139.2 | 140.7 | 143.9 | 150.3 | 163.1 |
| Q5_K_L | 132.58 GiB | 134.5 | 135.3 | 136.9 | 140.1 | 146.5 | 159.2 |
| Q5_K_M | 132.28 GiB | 134.2 | 135.0 | 136.6 | 139.8 | 146.2 | 158.9 |
| UD-Q5_K_S | 128.59 GiB | 130.5 | 131.3 | 132.9 | 136.1 | 142.5 | 155.2 |
| Q5_K_S | 128.01 GiB | 129.9 | 130.7 | 132.3 | 135.5 | 141.9 | 154.6 |
| Q4_1 | 116.67 GiB | 118.6 | 119.4 | 121.0 | 124.2 | 130.6 | 143.3 |
| UD-Q4_K_M | 113.71 GiB | 115.6 | 116.4 | 118.0 | 121.2 | 127.6 | 140.3 |
| Q4_K_L | 113.63 GiB | 115.6 | 116.4 | 117.9 | 121.1 | 127.5 | 140.3 |
| Q4_K_M | 113.27 GiB | 115.2 | 116.0 | 117.6 | 120.8 | 127.1 | 139.9 |
| Q4_K_S | 109.05 GiB | 111.0 | 111.8 | 113.4 | 116.5 | 122.9 | 135.7 |
| UD-Q4_K_S | 106.32 GiB | 108.2 | 109.0 | 110.6 | 113.8 | 120.2 | 132.9 |
| Q4_0 | 105.79 GiB | 107.7 | 108.5 | 110.1 | 113.3 | 119.7 | 132.4 |
| IQ4_NL | 105.54 GiB | 107.5 | 108.3 | 109.9 | 113.0 | 119.4 | 132.2 |
| IQ4_XS | 99.94 GiB | 101.9 | 102.7 | 104.2 | 107.4 | 113.8 | 126.6 |
| UD-IQ4_NL | 90.63 GiB | 92.6 | 93.4 | 94.9 | 98.1 | 104.5 | 117.3 |
| IQ3_M | 89.38 GiB | 91.3 | 92.1 | 93.7 | 96.9 | 103.3 | 116.0 |
| Q3_K_L | 88.94 GiB | 90.9 | 91.7 | 93.3 | 96.4 | 102.8 | 115.6 |
| UD-IQ4_XS | 88.79 GiB | 90.7 | 91.5 | 93.1 | 96.3 | 102.7 | 115.4 |
| Q3_K_M | 87.36 GiB | 89.3 | 90.1 | 91.7 | 94.9 | 101.2 | 114.0 |
| IQ3_XS | 85.44 GiB | 87.4 | 88.2 | 89.7 | 92.9 | 99.3 | 112.1 |
| UD-Q3_K_M | 83.13 GiB | 85.1 | 85.9 | 87.4 | 90.6 | 97.0 | 109.8 |
| Q3_K_S | 81.55 GiB | 83.5 | 84.3 | 85.9 | 89.1 | 95.4 | 108.2 |
| IQ3_XXS | 78.38 GiB | 80.3 | 81.1 | 82.7 | 85.9 | 92.3 | 105.0 |
| UD-IQ3_S | 74.54 GiB | 76.5 | 77.3 | 78.8 | 82.0 | 88.4 | 101.2 |
| UD-IQ3_XXS | 68.54 GiB | 70.5 | 71.3 | 72.9 | 76.0 | 82.4 | 95.2 |
| Q2_K_L | 66.82 GiB | 68.7 | 69.5 | 71.1 | 74.3 | 80.7 | 93.4 |
| Q2_K | 66.34 GiB | 68.3 | 69.1 | 70.6 | 73.8 | 80.2 | 93.0 |
| IQ2_M | 63.68 GiB | 65.6 | 66.4 | 68.0 | 71.2 | 77.6 | 90.3 |
| IQ2_S | 57.93 GiB | 59.8 | 60.6 | 62.2 | 65.4 | 71.8 | 84.6 |
| UD-IQ2_M | 57.58 GiB | 59.5 | 60.3 | 61.9 | 65.1 | 71.5 | 84.2 |
| UD-IQ2_XXS | 57.47 GiB | 59.4 | 60.2 | 61.8 | 65.0 | 71.3 | 84.1 |
| IQ2_XS | 56.76 GiB | 58.7 | 59.5 | 61.1 | 64.3 | 70.6 | 83.4 |
| UD-IQ1_M | 52.85 GiB | 54.8 | 55.6 | 57.2 | 60.4 | 66.7 | 79.5 |
| IQ2_XXS | 51.22 GiB | 53.1 | 53.9 | 55.5 | 58.7 | 65.1 | 77.8 |
| IQ1_M | 44.41 GiB | 46.3 | 47.1 | 48.7 | 51.9 | 58.3 | 71.0 |
| IQ1_S | 40.03 GiB | 42.0 | 42.8 | 44.3 | 47.5 | 53.9 | 66.7 |
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.