Does Step-3.5-Flash-REAP-121B-A11B fit in 24GB of VRAM?
Not at these settings. No indexed quantization of Step-3.5-Flash-REAP-121B-A11B fits 24GB card at any context we compute, with q4_0 KV. The smallest shipped quantization is 22.74 GiB in weights alone, against 22.32 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 | 94.1 | 94.5 | 95.3 | 97.0 | 100.4 | 107.1 |
| I1-Q5_K_M | 79.79 GiB | 81.3 | 81.8 | 82.6 | 84.3 | 87.7 | 94.4 |
| I1-Q5_K_S | 77.62 GiB | 79.2 | 79.6 | 80.4 | 82.1 | 85.5 | 92.2 |
| I1-Q4_1 | 70.61 GiB | 72.1 | 72.6 | 73.4 | 75.1 | 78.5 | 85.2 |
| I1-Q4_K_M | 67.82 GiB | 69.4 | 69.8 | 70.6 | 72.3 | 75.7 | 82.4 |
| I1-Q4_K_S | 63.83 GiB | 65.4 | 65.8 | 66.6 | 68.3 | 71.7 | 78.4 |
| I1-Q4_0 | 63.71 GiB | 65.3 | 65.7 | 66.5 | 68.2 | 71.6 | 78.3 |
| I1-IQ4_XS | 60.12 GiB | 61.7 | 62.1 | 62.9 | 64.6 | 68.0 | 74.7 |
| I1-Q3_K_L | 58.48 GiB | 60.0 | 60.4 | 61.3 | 63.0 | 66.3 | 73.1 |
| I1-Q3_K_M | 53.75 GiB | 55.3 | 55.7 | 56.6 | 58.2 | 61.6 | 68.4 |
| I1-IQ3_M | 49.23 GiB | 50.8 | 51.2 | 52.0 | 53.7 | 57.1 | 63.8 |
| I1-IQ3_S | 48.73 GiB | 50.3 | 50.7 | 51.5 | 53.2 | 56.6 | 63.3 |
| I1-Q3_K_S | 48.71 GiB | 50.2 | 50.7 | 51.5 | 53.2 | 56.6 | 63.3 |
| I1-IQ3_XS | 45.92 GiB | 47.5 | 47.9 | 48.7 | 50.4 | 53.8 | 60.5 |
| I1-IQ3_XXS | 43.40 GiB | 44.9 | 45.4 | 46.2 | 47.9 | 51.3 | 58.0 |
| I1-Q2_K | 41.19 GiB | 42.7 | 43.2 | 44.0 | 45.7 | 49.1 | 55.8 |
| I1-Q2_K_S | 37.79 GiB | 39.3 | 39.7 | 40.6 | 42.3 | 45.7 | 52.4 |
| I1-IQ2_M | 36.79 GiB | 38.3 | 38.8 | 39.6 | 41.3 | 44.7 | 51.4 |
| I1-IQ2_S | 33.40 GiB | 34.9 | 35.4 | 36.2 | 37.9 | 41.3 | 48.0 |
| I1-IQ2_XS | 32.98 GiB | 34.5 | 34.9 | 35.8 | 37.5 | 40.8 | 47.6 |
| I1-IQ2_XXS | 29.52 GiB | 31.1 | 31.5 | 32.3 | 34.0 | 37.4 | 44.1 |
| I1-IQ1_M | 25.28 GiB | 26.8 | 27.2 | 28.1 | 29.8 | 33.2 | 39.9 |
| I1-IQ1_S | 22.74 GiB | 24.3 | 24.7 | 25.5 | 27.2 | 30.6 | 37.4 |
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 there are no speeds on this page
A capacity is not a card. Whether a model fits depends only on memory, so every figure above holds for any 24GB accelerator. How fast it runs depends on memory bandwidth, which varies several-fold between cards of the same capacity — so putting a tokens-per-second number here would be inventing one. Pick a specific card from hardware and the speed column appears.
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.