Can I run Step-3.5-Flash-REAP-121B-A11B on a GeForce RTX 5070?
Not at these settings. No indexed quantization of Step-3.5-Flash-REAP-121B-A11B fits GeForce RTX 5070 at any context we compute, with q8_0 KV. The smallest shipped quantization is 22.74 GiB in weights alone, against 11.16 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 | 94.7 | 95.5 | 97.1 | 100.3 | 106.6 | 119.4 |
| I1-Q5_K_M | 79.79 GiB | 82.0 | 82.8 | 84.4 | 87.5 | 93.9 | 106.7 |
| I1-Q5_K_S | 77.62 GiB | 79.8 | 80.6 | 82.2 | 85.4 | 91.7 | 104.5 |
| I1-Q4_1 | 70.61 GiB | 72.8 | 73.6 | 75.2 | 78.4 | 84.7 | 97.5 |
| I1-Q4_K_M | 67.82 GiB | 70.0 | 70.8 | 72.4 | 75.6 | 81.9 | 94.7 |
| I1-Q4_K_S | 63.83 GiB | 66.0 | 66.8 | 68.4 | 71.6 | 78.0 | 90.7 |
| I1-Q4_0 | 63.71 GiB | 65.9 | 66.7 | 68.3 | 71.5 | 77.8 | 90.6 |
| I1-IQ4_XS | 60.12 GiB | 62.3 | 63.1 | 64.7 | 67.9 | 74.2 | 87.0 |
| I1-Q3_K_L | 58.48 GiB | 60.7 | 61.4 | 63.0 | 66.2 | 72.6 | 85.4 |
| I1-Q3_K_M | 53.75 GiB | 55.9 | 56.7 | 58.3 | 61.5 | 67.9 | 80.6 |
| I1-IQ3_M | 49.23 GiB | 51.4 | 52.2 | 53.8 | 57.0 | 63.4 | 76.1 |
| I1-IQ3_S | 48.73 GiB | 50.9 | 51.7 | 53.3 | 56.5 | 62.9 | 75.6 |
| I1-Q3_K_S | 48.71 GiB | 50.9 | 51.7 | 53.3 | 56.5 | 62.8 | 75.6 |
| I1-IQ3_XS | 45.92 GiB | 48.1 | 48.9 | 50.5 | 53.7 | 60.0 | 72.8 |
| I1-IQ3_XXS | 43.40 GiB | 45.6 | 46.4 | 48.0 | 51.1 | 57.5 | 70.3 |
| I1-Q2_K | 41.19 GiB | 43.4 | 44.2 | 45.8 | 48.9 | 55.3 | 68.1 |
| I1-Q2_K_S | 37.79 GiB | 40.0 | 40.8 | 42.3 | 45.5 | 51.9 | 64.7 |
| I1-IQ2_M | 36.79 GiB | 39.0 | 39.8 | 41.4 | 44.5 | 50.9 | 63.7 |
| I1-IQ2_S | 33.40 GiB | 35.6 | 36.4 | 38.0 | 41.1 | 47.5 | 60.3 |
| I1-IQ2_XS | 32.98 GiB | 35.1 | 35.9 | 37.5 | 40.7 | 47.1 | 59.9 |
| I1-IQ2_XXS | 29.52 GiB | 31.7 | 32.5 | 34.1 | 37.3 | 43.6 | 56.4 |
| I1-IQ1_M | 25.28 GiB | 27.5 | 28.3 | 29.8 | 33.0 | 39.4 | 52.2 |
| I1-IQ1_S | 22.74 GiB | 24.9 | 25.7 | 27.3 | 30.5 | 36.9 | 49.6 |
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.