Can I run Step-3.7-Flash on a Radeon RX 6500 XT?
Not at these settings. No indexed quantization of Step-3.7-Flash fits Radeon RX 6500 XT at any context we compute, with q4_0 KV. The smallest shipped quantization is 40.03 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◐ |
|---|---|---|---|---|---|---|---|
| BF16 | 366.96 GiB | 368.6 | 369.0 | 369.9 | 371.6 | 374.9 | 381.7 |
| Q8_0 | 197.44 GiB | 199.1 | 199.5 | 200.3 | 202.0 | 205.4 | 212.2 |
| Q6_K_L | 160.20 GiB | 161.8 | 162.3 | 163.1 | 164.8 | 168.2 | 174.9 |
| Q6_K | 159.96 GiB | 161.6 | 162.0 | 162.9 | 164.6 | 167.9 | 174.7 |
| UD-Q6_K | 152.12 GiB | 153.8 | 154.2 | 155.0 | 156.7 | 160.1 | 166.8 |
| UD-Q5_K_M | 136.43 GiB | 138.1 | 138.5 | 139.3 | 141.0 | 144.4 | 151.2 |
| Q5_K_L | 132.58 GiB | 134.2 | 134.6 | 135.5 | 137.2 | 140.6 | 147.3 |
| Q5_K_M | 132.28 GiB | 133.9 | 134.3 | 135.2 | 136.9 | 140.2 | 147.0 |
| UD-Q5_K_S | 128.59 GiB | 130.2 | 130.7 | 131.5 | 133.2 | 136.6 | 143.3 |
| Q5_K_S | 128.01 GiB | 129.7 | 130.1 | 130.9 | 132.6 | 136.0 | 142.7 |
| Q4_1 | 116.67 GiB | 118.3 | 118.7 | 119.6 | 121.3 | 124.6 | 131.4 |
| UD-Q4_K_M | 113.71 GiB | 115.3 | 115.8 | 116.6 | 118.3 | 121.7 | 128.4 |
| Q4_K_L | 113.63 GiB | 115.3 | 115.7 | 116.5 | 118.2 | 121.6 | 128.4 |
| Q4_K_M | 113.27 GiB | 114.9 | 115.3 | 116.2 | 117.9 | 121.2 | 128.0 |
| Q4_K_S | 109.05 GiB | 110.7 | 111.1 | 112.0 | 113.6 | 117.0 | 123.8 |
| UD-Q4_K_S | 106.32 GiB | 108.0 | 108.4 | 109.2 | 110.9 | 114.3 | 121.0 |
| Q4_0 | 105.79 GiB | 107.4 | 107.9 | 108.7 | 110.4 | 113.8 | 120.5 |
| IQ4_NL | 105.54 GiB | 107.2 | 107.6 | 108.4 | 110.1 | 113.5 | 120.3 |
| IQ4_XS | 99.94 GiB | 101.6 | 102.0 | 102.8 | 104.5 | 107.9 | 114.7 |
| UD-IQ4_NL | 90.63 GiB | 92.3 | 92.7 | 93.5 | 95.2 | 98.6 | 105.4 |
| IQ3_M | 89.38 GiB | 91.0 | 91.4 | 92.3 | 94.0 | 97.4 | 104.1 |
| Q3_K_L | 88.94 GiB | 90.6 | 91.0 | 91.8 | 93.5 | 96.9 | 103.7 |
| UD-IQ4_XS | 88.79 GiB | 90.4 | 90.9 | 91.7 | 93.4 | 96.8 | 103.5 |
| Q3_K_M | 87.36 GiB | 89.0 | 89.4 | 90.3 | 92.0 | 95.3 | 102.1 |
| IQ3_XS | 85.44 GiB | 87.1 | 87.5 | 88.3 | 90.0 | 93.4 | 100.2 |
| UD-Q3_K_M | 83.13 GiB | 84.8 | 85.2 | 86.0 | 87.7 | 91.1 | 97.9 |
| Q3_K_S | 81.55 GiB | 83.2 | 83.6 | 84.5 | 86.1 | 89.5 | 96.3 |
| IQ3_XXS | 78.38 GiB | 80.0 | 80.4 | 81.3 | 83.0 | 86.4 | 93.1 |
| UD-IQ3_S | 74.54 GiB | 76.2 | 76.6 | 77.4 | 79.1 | 82.5 | 89.3 |
| UD-IQ3_XXS | 68.54 GiB | 70.2 | 70.6 | 71.4 | 73.1 | 76.5 | 83.3 |
| Q2_K_L | 66.82 GiB | 68.5 | 68.9 | 69.7 | 71.4 | 74.8 | 81.5 |
| Q2_K | 66.34 GiB | 68.0 | 68.4 | 69.2 | 70.9 | 74.3 | 81.1 |
| IQ2_M | 63.68 GiB | 65.3 | 65.7 | 66.6 | 68.3 | 71.7 | 78.4 |
| IQ2_S | 57.93 GiB | 59.6 | 60.0 | 60.8 | 62.5 | 65.9 | 72.6 |
| UD-IQ2_M | 57.58 GiB | 59.2 | 59.6 | 60.5 | 62.2 | 65.5 | 72.3 |
| UD-IQ2_XXS | 57.47 GiB | 59.1 | 59.5 | 60.4 | 62.1 | 65.4 | 72.2 |
| IQ2_XS | 56.76 GiB | 58.4 | 58.8 | 59.7 | 61.3 | 64.7 | 71.5 |
| UD-IQ1_M | 52.85 GiB | 54.5 | 54.9 | 55.8 | 57.4 | 60.8 | 67.6 |
| IQ2_XXS | 51.22 GiB | 52.9 | 53.3 | 54.1 | 55.8 | 59.2 | 65.9 |
| IQ1_M | 44.41 GiB | 46.0 | 46.5 | 47.3 | 49.0 | 52.4 | 59.1 |
| IQ1_S | 40.03 GiB | 41.7 | 42.1 | 42.9 | 44.6 | 48.0 | 54.8 |
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.