Can I run step-3.5-flash on a GeForce RTX 5090 D?
Not at these settings. No indexed quantization of step-3.5-flash fits GeForce RTX 5090 D at any context we compute, with q8_0 KV. The smallest shipped quantization is 38.47 GiB in weights alone, against 29.76 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 | 197.2 | 198.0 | 199.6 | 202.8 | 209.2 | 221.9 |
| Q6_K | 150.81 GiB | 153.0 | 153.8 | 155.4 | 158.6 | 164.9 | 177.7 |
| Q5_K_L | 130.88 GiB | 133.1 | 133.9 | 135.4 | 138.6 | 145.0 | 157.8 |
| Q5_K_M | 130.58 GiB | 132.8 | 133.5 | 135.1 | 138.3 | 144.7 | 157.5 |
| Q5_K_S | 126.55 GiB | 128.7 | 129.5 | 131.1 | 134.3 | 140.7 | 153.4 |
| Q4_1 | 115.15 GiB | 117.3 | 118.1 | 119.7 | 122.9 | 129.3 | 142.0 |
| Q4_K_L | 112.07 GiB | 114.2 | 115.0 | 116.6 | 119.8 | 126.2 | 138.9 |
| Q4_K_M | 111.70 GiB | 113.9 | 114.7 | 116.3 | 119.5 | 125.8 | 138.6 |
| Q4_K | 110.56 GiB | 112.7 | 113.5 | 115.1 | 118.3 | 124.7 | 137.4 |
| Q4_K_S | 107.70 GiB | 109.9 | 110.7 | 112.3 | 115.5 | 121.8 | 134.6 |
| Q4_0 | 105.60 GiB | 107.8 | 108.6 | 110.2 | 113.4 | 119.7 | 132.5 |
| IQ4_NL | 103.97 GiB | 106.1 | 106.9 | 108.5 | 111.7 | 118.1 | 130.9 |
| IQ4_XS | 98.34 GiB | 100.5 | 101.3 | 102.9 | 106.1 | 112.5 | 125.2 |
| Q3_K_L | 87.26 GiB | 89.4 | 90.2 | 91.8 | 95.0 | 101.4 | 114.1 |
| Q3_K_M | 84.04 GiB | 86.2 | 87.0 | 88.6 | 91.8 | 98.2 | 110.9 |
| IQ3_M | 83.98 GiB | 86.1 | 86.9 | 88.5 | 91.7 | 98.1 | 110.9 |
| Q3_K_S | 80.06 GiB | 82.2 | 83.0 | 84.6 | 87.8 | 94.2 | 106.9 |
| IQ3_XS | 75.46 GiB | 77.6 | 78.4 | 80.0 | 83.2 | 89.6 | 102.3 |
| IQ3_XXS | 72.82 GiB | 75.0 | 75.8 | 77.4 | 80.6 | 86.9 | 99.7 |
| Q2_K_L | 65.26 GiB | 67.4 | 68.2 | 69.8 | 73.0 | 79.4 | 92.1 |
| Q2_K | 64.78 GiB | 66.9 | 67.7 | 69.3 | 72.5 | 78.9 | 91.7 |
| IQ2_M | 59.59 GiB | 61.8 | 62.6 | 64.1 | 67.3 | 73.7 | 86.5 |
| IQ2_S | 51.70 GiB | 53.9 | 54.7 | 56.3 | 59.5 | 65.8 | 78.6 |
| IQ2_XS | 51.40 GiB | 53.6 | 54.4 | 56.0 | 59.1 | 65.5 | 78.3 |
| IQ2_XXS | 44.74 GiB | 46.9 | 47.7 | 49.3 | 52.5 | 58.9 | 71.6 |
| IQ1_M | 40.17 GiB | 42.3 | 43.1 | 44.7 | 47.9 | 54.3 | 67.0 |
| IQ1_S | 38.47 GiB | 40.6 | 41.4 | 43.0 | 46.2 | 52.6 | 65.3 |
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.