Can I run step-3.5-flash on a Apple M5 Max?
Not at these settings. No indexed quantization of step-3.5-flash fits Apple M5 Max at any context we compute, with q4_0 KV. The smallest shipped quantization is 38.47 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◐ |
|---|---|---|---|---|---|---|---|
| Q8_0 | 195.04 GiB | 196.3 | 196.7 | 197.6 | 199.3 | 202.7 | 209.4 |
| Q6_K | 150.81 GiB | 152.1 | 152.5 | 153.4 | 155.1 | 158.4 | 165.2 |
| Q5_K_L | 130.88 GiB | 132.2 | 132.6 | 133.4 | 135.1 | 138.5 | 145.3 |
| Q5_K_M | 130.58 GiB | 131.9 | 132.3 | 133.1 | 134.8 | 138.2 | 144.9 |
| Q5_K_S | 126.55 GiB | 127.8 | 128.3 | 129.1 | 130.8 | 134.2 | 140.9 |
| Q4_1 | 115.15 GiB | 116.4 | 116.9 | 117.7 | 119.4 | 122.8 | 129.5 |
| Q4_K_L | 112.07 GiB | 113.4 | 113.8 | 114.6 | 116.3 | 119.7 | 126.4 |
| Q4_K_M | 111.70 GiB | 113.0 | 113.4 | 114.3 | 115.9 | 119.3 | 126.1 |
| Q4_K | 110.56 GiB | 111.8 | 112.3 | 113.1 | 114.8 | 118.2 | 124.9 |
| Q4_K_S | 107.70 GiB | 109.0 | 109.4 | 110.3 | 111.9 | 115.3 | 122.1 |
| Q4_0 | 105.60 GiB | 106.9 | 107.3 | 108.2 | 109.8 | 113.2 | 120.0 |
| IQ4_NL | 103.97 GiB | 105.3 | 105.7 | 106.5 | 108.2 | 111.6 | 118.3 |
| IQ4_XS | 98.34 GiB | 99.6 | 100.1 | 100.9 | 102.6 | 106.0 | 112.7 |
| Q3_K_L | 87.26 GiB | 88.5 | 89.0 | 89.8 | 91.5 | 94.9 | 101.6 |
| Q3_K_M | 84.04 GiB | 85.3 | 85.7 | 86.6 | 88.3 | 91.7 | 98.4 |
| IQ3_M | 83.98 GiB | 85.3 | 85.7 | 86.5 | 88.2 | 91.6 | 98.3 |
| Q3_K_S | 80.06 GiB | 81.4 | 81.8 | 82.6 | 84.3 | 87.7 | 94.4 |
| IQ3_XS | 75.46 GiB | 76.8 | 77.2 | 78.0 | 79.7 | 83.1 | 89.8 |
| IQ3_XXS | 72.82 GiB | 74.1 | 74.5 | 75.4 | 77.1 | 80.4 | 87.2 |
| Q2_K_L | 65.26 GiB | 66.5 | 67.0 | 67.8 | 69.5 | 72.9 | 79.6 |
| Q2_K | 64.78 GiB | 66.1 | 66.5 | 67.3 | 69.0 | 72.4 | 79.1 |
| IQ2_M | 59.59 GiB | 60.9 | 61.3 | 62.1 | 63.8 | 67.2 | 74.0 |
| IQ2_S | 51.70 GiB | 53.0 | 53.4 | 54.3 | 55.9 | 59.3 | 66.1 |
| IQ2_XS | 51.40 GiB | 52.7 | 53.1 | 54.0 | 55.6 | 59.0 | 65.8 |
| IQ2_XXS | 44.74 GiB | 46.0 | 46.4 | 47.3 | 49.0 | 52.4 | 59.1 |
| IQ1_M | 40.17 GiB | 41.5 | 41.9 | 42.7 | 44.4 | 47.8 | 54.5 |
| IQ1_S | 38.47 GiB | 39.8 | 40.2 | 41.0 | 42.7 | 46.1 | 52.8 |
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.