Can I run Step-3.5-Flash-REAP-121B-A11B on a Radeon RX 6500 XT?
Not at these settings. No indexed quantization of Step-3.5-Flash-REAP-121B-A11B fits Radeon RX 6500 XT at any context we compute, with q8_0 KV. The smallest shipped quantization is 22.74 GiB in weights alone, against 3.72 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 | 94.8 | 95.6 | 97.2 | 100.4 | 106.7 | 119.5 |
| I1-Q5_K_M | 79.79 GiB | 82.1 | 82.9 | 84.5 | 87.6 | 94.0 | 106.8 |
| I1-Q5_K_S | 77.62 GiB | 79.9 | 80.7 | 82.3 | 85.5 | 91.8 | 104.6 |
| I1-Q4_1 | 70.61 GiB | 72.9 | 73.7 | 75.3 | 78.5 | 84.8 | 97.6 |
| I1-Q4_K_M | 67.82 GiB | 70.1 | 70.9 | 72.5 | 75.7 | 82.0 | 94.8 |
| I1-Q4_K_S | 63.83 GiB | 66.1 | 66.9 | 68.5 | 71.7 | 78.1 | 90.8 |
| I1-Q4_0 | 63.71 GiB | 66.0 | 66.8 | 68.4 | 71.6 | 77.9 | 90.7 |
| I1-IQ4_XS | 60.12 GiB | 62.4 | 63.2 | 64.8 | 68.0 | 74.3 | 87.1 |
| I1-Q3_K_L | 58.48 GiB | 60.8 | 61.5 | 63.1 | 66.3 | 72.7 | 85.5 |
| I1-Q3_K_M | 53.75 GiB | 56.0 | 56.8 | 58.4 | 61.6 | 68.0 | 80.7 |
| I1-IQ3_M | 49.23 GiB | 51.5 | 52.3 | 53.9 | 57.1 | 63.5 | 76.2 |
| I1-IQ3_S | 48.73 GiB | 51.0 | 51.8 | 53.4 | 56.6 | 63.0 | 75.7 |
| I1-Q3_K_S | 48.71 GiB | 51.0 | 51.8 | 53.4 | 56.6 | 62.9 | 75.7 |
| I1-IQ3_XS | 45.92 GiB | 48.2 | 49.0 | 50.6 | 53.8 | 60.1 | 72.9 |
| I1-IQ3_XXS | 43.40 GiB | 45.7 | 46.5 | 48.1 | 51.2 | 57.6 | 70.4 |
| I1-Q2_K | 41.19 GiB | 43.5 | 44.3 | 45.9 | 49.0 | 55.4 | 68.2 |
| I1-Q2_K_S | 37.79 GiB | 40.1 | 40.9 | 42.4 | 45.6 | 52.0 | 64.8 |
| I1-IQ2_M | 36.79 GiB | 39.1 | 39.9 | 41.5 | 44.6 | 51.0 | 63.8 |
| I1-IQ2_S | 33.40 GiB | 35.7 | 36.5 | 38.1 | 41.2 | 47.6 | 60.4 |
| I1-IQ2_XS | 32.98 GiB | 35.2 | 36.0 | 37.6 | 40.8 | 47.2 | 60.0 |
| I1-IQ2_XXS | 29.52 GiB | 31.8 | 32.6 | 34.2 | 37.4 | 43.7 | 56.5 |
| I1-IQ1_M | 25.28 GiB | 27.6 | 28.4 | 29.9 | 33.1 | 39.5 | 52.3 |
| I1-IQ1_S | 22.74 GiB | 25.0 | 25.8 | 27.4 | 30.6 | 37.0 | 49.7 |
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.