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 q4_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 | 94.2 | 94.6 | 95.4 | 97.1 | 100.5 | 107.2 |
| I1-Q5_K_M | 79.79 GiB | 81.4 | 81.9 | 82.7 | 84.4 | 87.8 | 94.5 |
| I1-Q5_K_S | 77.62 GiB | 79.3 | 79.7 | 80.5 | 82.2 | 85.6 | 92.3 |
| I1-Q4_1 | 70.61 GiB | 72.2 | 72.7 | 73.5 | 75.2 | 78.6 | 85.3 |
| I1-Q4_K_M | 67.82 GiB | 69.5 | 69.9 | 70.7 | 72.4 | 75.8 | 82.5 |
| I1-Q4_K_S | 63.83 GiB | 65.5 | 65.9 | 66.7 | 68.4 | 71.8 | 78.5 |
| I1-Q4_0 | 63.71 GiB | 65.4 | 65.8 | 66.6 | 68.3 | 71.7 | 78.4 |
| I1-IQ4_XS | 60.12 GiB | 61.8 | 62.2 | 63.0 | 64.7 | 68.1 | 74.8 |
| I1-Q3_K_L | 58.48 GiB | 60.1 | 60.5 | 61.4 | 63.1 | 66.4 | 73.2 |
| I1-Q3_K_M | 53.75 GiB | 55.4 | 55.8 | 56.7 | 58.3 | 61.7 | 68.5 |
| I1-IQ3_M | 49.23 GiB | 50.9 | 51.3 | 52.1 | 53.8 | 57.2 | 63.9 |
| I1-IQ3_S | 48.73 GiB | 50.4 | 50.8 | 51.6 | 53.3 | 56.7 | 63.4 |
| I1-Q3_K_S | 48.71 GiB | 50.3 | 50.8 | 51.6 | 53.3 | 56.7 | 63.4 |
| I1-IQ3_XS | 45.92 GiB | 47.6 | 48.0 | 48.8 | 50.5 | 53.9 | 60.6 |
| I1-IQ3_XXS | 43.40 GiB | 45.0 | 45.5 | 46.3 | 48.0 | 51.4 | 58.1 |
| I1-Q2_K | 41.19 GiB | 42.8 | 43.3 | 44.1 | 45.8 | 49.2 | 55.9 |
| I1-Q2_K_S | 37.79 GiB | 39.4 | 39.8 | 40.7 | 42.4 | 45.8 | 52.5 |
| I1-IQ2_M | 36.79 GiB | 38.4 | 38.9 | 39.7 | 41.4 | 44.8 | 51.5 |
| I1-IQ2_S | 33.40 GiB | 35.0 | 35.5 | 36.3 | 38.0 | 41.4 | 48.1 |
| I1-IQ2_XS | 32.98 GiB | 34.6 | 35.0 | 35.9 | 37.6 | 40.9 | 47.7 |
| I1-IQ2_XXS | 29.52 GiB | 31.2 | 31.6 | 32.4 | 34.1 | 37.5 | 44.2 |
| I1-IQ1_M | 25.28 GiB | 26.9 | 27.3 | 28.2 | 29.9 | 33.3 | 40.0 |
| I1-IQ1_S | 22.74 GiB | 24.4 | 24.8 | 25.6 | 27.3 | 30.7 | 37.5 |
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.