Can I run Step-3.7-Flash on a GeForce RTX 5050?
Not at these settings. No indexed quantization of Step-3.7-Flash fits GeForce RTX 5050 at any context we compute, with q8_0 KV. The smallest shipped quantization is 40.03 GiB in weights alone, against 7.44 GiB usable. CPU offload can still run it, slowly.
Every quantization at every context
| Quant | Weights● | 4K◐ | 8K◐ | 16K◐ | 32K◐ | 64K◐ | 128K◐ |
|---|---|---|---|---|---|---|---|
| BF16 | 366.96 GiB | 369.1 | 369.9 | 371.5 | 374.7 | 381.1 | 393.8 |
| Q8_0 | 197.44 GiB | 199.6 | 200.4 | 202.0 | 205.2 | 211.6 | 224.3 |
| Q6_K_L | 160.20 GiB | 162.4 | 163.2 | 164.8 | 168.0 | 174.3 | 187.1 |
| Q6_K | 159.96 GiB | 162.1 | 162.9 | 164.5 | 167.7 | 174.1 | 186.8 |
| UD-Q6_K | 152.12 GiB | 154.3 | 155.1 | 156.7 | 159.9 | 166.2 | 179.0 |
| UD-Q5_K_M | 136.43 GiB | 138.6 | 139.4 | 141.0 | 144.2 | 150.6 | 163.3 |
| Q5_K_L | 132.58 GiB | 134.8 | 135.6 | 137.1 | 140.3 | 146.7 | 159.5 |
| Q5_K_M | 132.28 GiB | 134.5 | 135.2 | 136.8 | 140.0 | 146.4 | 159.2 |
| UD-Q5_K_S | 128.59 GiB | 130.8 | 131.6 | 133.2 | 136.3 | 142.7 | 155.5 |
| Q5_K_S | 128.01 GiB | 130.2 | 131.0 | 132.6 | 135.8 | 142.1 | 154.9 |
| Q4_1 | 116.67 GiB | 118.8 | 119.6 | 121.2 | 124.4 | 130.8 | 143.6 |
| UD-Q4_K_M | 113.71 GiB | 115.9 | 116.7 | 118.3 | 121.5 | 127.8 | 140.6 |
| Q4_K_L | 113.63 GiB | 115.8 | 116.6 | 118.2 | 121.4 | 127.8 | 140.5 |
| Q4_K_M | 113.27 GiB | 115.4 | 116.2 | 117.8 | 121.0 | 127.4 | 140.1 |
| Q4_K_S | 109.05 GiB | 111.2 | 112.0 | 113.6 | 116.8 | 123.2 | 135.9 |
| UD-Q4_K_S | 106.32 GiB | 108.5 | 109.3 | 110.9 | 114.1 | 120.4 | 133.2 |
| Q4_0 | 105.79 GiB | 108.0 | 108.8 | 110.4 | 113.5 | 119.9 | 132.7 |
| IQ4_NL | 105.54 GiB | 107.7 | 108.5 | 110.1 | 113.3 | 119.7 | 132.4 |
| IQ4_XS | 99.94 GiB | 102.1 | 102.9 | 104.5 | 107.7 | 114.1 | 126.8 |
| UD-IQ4_NL | 90.63 GiB | 92.8 | 93.6 | 95.2 | 98.4 | 104.8 | 117.5 |
| IQ3_M | 89.38 GiB | 91.6 | 92.4 | 93.9 | 97.1 | 103.5 | 116.3 |
| Q3_K_L | 88.94 GiB | 91.1 | 91.9 | 93.5 | 96.7 | 103.1 | 115.8 |
| UD-IQ4_XS | 88.79 GiB | 91.0 | 91.8 | 93.4 | 96.5 | 102.9 | 115.7 |
| Q3_K_M | 87.36 GiB | 89.5 | 90.3 | 91.9 | 95.1 | 101.5 | 114.2 |
| IQ3_XS | 85.44 GiB | 87.6 | 88.4 | 90.0 | 93.2 | 99.6 | 112.3 |
| UD-Q3_K_M | 83.13 GiB | 85.3 | 86.1 | 87.7 | 90.9 | 97.3 | 110.0 |
| Q3_K_S | 81.55 GiB | 83.7 | 84.5 | 86.1 | 89.3 | 95.7 | 108.4 |
| IQ3_XXS | 78.38 GiB | 80.6 | 81.4 | 82.9 | 86.1 | 92.5 | 105.3 |
| UD-IQ3_S | 74.54 GiB | 76.7 | 77.5 | 79.1 | 82.3 | 88.7 | 101.4 |
| UD-IQ3_XXS | 68.54 GiB | 70.7 | 71.5 | 73.1 | 76.3 | 82.7 | 95.4 |
| Q2_K_L | 66.82 GiB | 69.0 | 69.8 | 71.4 | 74.6 | 80.9 | 93.7 |
| Q2_K | 66.34 GiB | 68.5 | 69.3 | 70.9 | 74.1 | 80.5 | 93.2 |
| IQ2_M | 63.68 GiB | 65.9 | 66.7 | 68.2 | 71.4 | 77.8 | 90.6 |
| IQ2_S | 57.93 GiB | 60.1 | 60.9 | 62.5 | 65.7 | 72.1 | 84.8 |
| UD-IQ2_M | 57.58 GiB | 59.7 | 60.5 | 62.1 | 65.3 | 71.7 | 84.5 |
| UD-IQ2_XXS | 57.47 GiB | 59.6 | 60.4 | 62.0 | 65.2 | 71.6 | 84.3 |
| IQ2_XS | 56.76 GiB | 58.9 | 59.7 | 61.3 | 64.5 | 70.9 | 83.6 |
| UD-IQ1_M | 52.85 GiB | 55.0 | 55.8 | 57.4 | 60.6 | 67.0 | 79.7 |
| IQ2_XXS | 51.22 GiB | 53.4 | 54.2 | 55.8 | 59.0 | 65.3 | 78.1 |
| IQ1_M | 44.41 GiB | 46.6 | 47.4 | 49.0 | 52.2 | 58.5 | 71.3 |
| IQ1_S | 40.03 GiB | 42.2 | 43.0 | 44.6 | 47.8 | 54.2 | 66.9 |
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.