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 q8_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 | 197.0 | 197.8 | 199.3 | 202.5 | 208.9 | 221.7 |
| Q6_K | 150.81 GiB | 152.7 | 153.5 | 155.1 | 158.3 | 164.7 | 177.4 |
| Q5_K_L | 130.88 GiB | 132.8 | 133.6 | 135.2 | 138.4 | 144.8 | 157.5 |
| Q5_K_M | 130.58 GiB | 132.5 | 133.3 | 134.9 | 138.1 | 144.5 | 157.2 |
| Q5_K_S | 126.55 GiB | 128.5 | 129.3 | 130.9 | 134.1 | 140.4 | 153.2 |
| Q4_1 | 115.15 GiB | 117.1 | 117.9 | 119.5 | 122.7 | 129.0 | 141.8 |
| Q4_K_L | 112.07 GiB | 114.0 | 114.8 | 116.4 | 119.6 | 125.9 | 138.7 |
| Q4_K_M | 111.70 GiB | 113.6 | 114.4 | 116.0 | 119.2 | 125.6 | 138.3 |
| Q4_K | 110.56 GiB | 112.5 | 113.3 | 114.9 | 118.1 | 124.4 | 137.2 |
| Q4_K_S | 107.70 GiB | 109.6 | 110.4 | 112.0 | 115.2 | 121.6 | 134.3 |
| Q4_0 | 105.60 GiB | 107.5 | 108.3 | 109.9 | 113.1 | 119.5 | 132.2 |
| IQ4_NL | 103.97 GiB | 105.9 | 106.7 | 108.3 | 111.5 | 117.9 | 130.6 |
| IQ4_XS | 98.34 GiB | 100.3 | 101.1 | 102.7 | 105.8 | 112.2 | 125.0 |
| Q3_K_L | 87.26 GiB | 89.2 | 90.0 | 91.6 | 94.8 | 101.1 | 113.9 |
| Q3_K_M | 84.04 GiB | 86.0 | 86.8 | 88.3 | 91.5 | 97.9 | 110.7 |
| IQ3_M | 83.98 GiB | 85.9 | 86.7 | 88.3 | 91.5 | 97.9 | 110.6 |
| Q3_K_S | 80.06 GiB | 82.0 | 82.8 | 84.4 | 87.6 | 93.9 | 106.7 |
| IQ3_XS | 75.46 GiB | 77.4 | 78.2 | 79.8 | 83.0 | 89.3 | 102.1 |
| IQ3_XXS | 72.82 GiB | 74.7 | 75.5 | 77.1 | 80.3 | 86.7 | 99.4 |
| Q2_K_L | 65.26 GiB | 67.2 | 68.0 | 69.6 | 72.8 | 79.1 | 91.9 |
| Q2_K | 64.78 GiB | 66.7 | 67.5 | 69.1 | 72.3 | 78.7 | 91.4 |
| IQ2_M | 59.59 GiB | 61.5 | 62.3 | 63.9 | 67.1 | 73.5 | 86.2 |
| IQ2_S | 51.70 GiB | 53.6 | 54.4 | 56.0 | 59.2 | 65.6 | 78.3 |
| IQ2_XS | 51.40 GiB | 53.3 | 54.1 | 55.7 | 58.9 | 65.3 | 78.0 |
| IQ2_XXS | 44.74 GiB | 46.7 | 47.5 | 49.0 | 52.2 | 58.6 | 71.4 |
| IQ1_M | 40.17 GiB | 42.1 | 42.9 | 44.5 | 47.7 | 54.0 | 66.8 |
| IQ1_S | 38.47 GiB | 40.4 | 41.2 | 42.8 | 46.0 | 52.3 | 65.1 |
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.