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 q8_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 | 369.2 | 370.0 | 371.6 | 374.8 | 381.2 | 393.9 |
| Q8_0 | 197.44 GiB | 199.7 | 200.5 | 202.1 | 205.3 | 211.7 | 224.4 |
| Q6_K_L | 160.20 GiB | 162.5 | 163.3 | 164.9 | 168.1 | 174.4 | 187.2 |
| Q6_K | 159.96 GiB | 162.2 | 163.0 | 164.6 | 167.8 | 174.2 | 186.9 |
| UD-Q6_K | 152.12 GiB | 154.4 | 155.2 | 156.8 | 160.0 | 166.3 | 179.1 |
| UD-Q5_K_M | 136.43 GiB | 138.7 | 139.5 | 141.1 | 144.3 | 150.7 | 163.4 |
| Q5_K_L | 132.58 GiB | 134.9 | 135.7 | 137.2 | 140.4 | 146.8 | 159.6 |
| Q5_K_M | 132.28 GiB | 134.6 | 135.3 | 136.9 | 140.1 | 146.5 | 159.3 |
| UD-Q5_K_S | 128.59 GiB | 130.9 | 131.7 | 133.3 | 136.4 | 142.8 | 155.6 |
| Q5_K_S | 128.01 GiB | 130.3 | 131.1 | 132.7 | 135.9 | 142.2 | 155.0 |
| Q4_1 | 116.67 GiB | 118.9 | 119.7 | 121.3 | 124.5 | 130.9 | 143.7 |
| UD-Q4_K_M | 113.71 GiB | 116.0 | 116.8 | 118.4 | 121.6 | 127.9 | 140.7 |
| Q4_K_L | 113.63 GiB | 115.9 | 116.7 | 118.3 | 121.5 | 127.9 | 140.6 |
| Q4_K_M | 113.27 GiB | 115.5 | 116.3 | 117.9 | 121.1 | 127.5 | 140.2 |
| Q4_K_S | 109.05 GiB | 111.3 | 112.1 | 113.7 | 116.9 | 123.3 | 136.0 |
| UD-Q4_K_S | 106.32 GiB | 108.6 | 109.4 | 111.0 | 114.2 | 120.5 | 133.3 |
| Q4_0 | 105.79 GiB | 108.1 | 108.9 | 110.5 | 113.6 | 120.0 | 132.8 |
| IQ4_NL | 105.54 GiB | 107.8 | 108.6 | 110.2 | 113.4 | 119.8 | 132.5 |
| IQ4_XS | 99.94 GiB | 102.2 | 103.0 | 104.6 | 107.8 | 114.2 | 126.9 |
| UD-IQ4_NL | 90.63 GiB | 92.9 | 93.7 | 95.3 | 98.5 | 104.9 | 117.6 |
| IQ3_M | 89.38 GiB | 91.7 | 92.5 | 94.0 | 97.2 | 103.6 | 116.4 |
| Q3_K_L | 88.94 GiB | 91.2 | 92.0 | 93.6 | 96.8 | 103.2 | 115.9 |
| UD-IQ4_XS | 88.79 GiB | 91.1 | 91.9 | 93.5 | 96.6 | 103.0 | 115.8 |
| Q3_K_M | 87.36 GiB | 89.6 | 90.4 | 92.0 | 95.2 | 101.6 | 114.3 |
| IQ3_XS | 85.44 GiB | 87.7 | 88.5 | 90.1 | 93.3 | 99.7 | 112.4 |
| UD-Q3_K_M | 83.13 GiB | 85.4 | 86.2 | 87.8 | 91.0 | 97.4 | 110.1 |
| Q3_K_S | 81.55 GiB | 83.8 | 84.6 | 86.2 | 89.4 | 95.8 | 108.5 |
| IQ3_XXS | 78.38 GiB | 80.7 | 81.5 | 83.0 | 86.2 | 92.6 | 105.4 |
| UD-IQ3_S | 74.54 GiB | 76.8 | 77.6 | 79.2 | 82.4 | 88.8 | 101.5 |
| UD-IQ3_XXS | 68.54 GiB | 70.8 | 71.6 | 73.2 | 76.4 | 82.8 | 95.5 |
| Q2_K_L | 66.82 GiB | 69.1 | 69.9 | 71.5 | 74.7 | 81.0 | 93.8 |
| Q2_K | 66.34 GiB | 68.6 | 69.4 | 71.0 | 74.2 | 80.6 | 93.3 |
| IQ2_M | 63.68 GiB | 66.0 | 66.8 | 68.3 | 71.5 | 77.9 | 90.7 |
| IQ2_S | 57.93 GiB | 60.2 | 61.0 | 62.6 | 65.8 | 72.2 | 84.9 |
| UD-IQ2_M | 57.58 GiB | 59.8 | 60.6 | 62.2 | 65.4 | 71.8 | 84.6 |
| UD-IQ2_XXS | 57.47 GiB | 59.7 | 60.5 | 62.1 | 65.3 | 71.7 | 84.4 |
| IQ2_XS | 56.76 GiB | 59.0 | 59.8 | 61.4 | 64.6 | 71.0 | 83.7 |
| UD-IQ1_M | 52.85 GiB | 55.1 | 55.9 | 57.5 | 60.7 | 67.1 | 79.8 |
| IQ2_XXS | 51.22 GiB | 53.5 | 54.3 | 55.9 | 59.1 | 65.4 | 78.2 |
| IQ1_M | 44.41 GiB | 46.7 | 47.5 | 49.1 | 52.3 | 58.6 | 71.4 |
| IQ1_S | 40.03 GiB | 42.3 | 43.1 | 44.7 | 47.9 | 54.3 | 67.0 |
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.