Can I run step-3.5-flash on a Radeon RX 6500 XT?
Not at these settings. No indexed quantization of step-3.5-flash fits Radeon RX 6500 XT at any context we compute, with q8_0 KV. The smallest shipped quantization is 38.47 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◐ |
|---|---|---|---|---|---|---|---|
| Q8_0 | 195.04 GiB | 197.3 | 198.1 | 199.7 | 202.9 | 209.3 | 222.0 |
| Q6_K | 150.81 GiB | 153.1 | 153.9 | 155.5 | 158.7 | 165.0 | 177.8 |
| Q5_K_L | 130.88 GiB | 133.2 | 134.0 | 135.5 | 138.7 | 145.1 | 157.9 |
| Q5_K_M | 130.58 GiB | 132.9 | 133.6 | 135.2 | 138.4 | 144.8 | 157.6 |
| Q5_K_S | 126.55 GiB | 128.8 | 129.6 | 131.2 | 134.4 | 140.8 | 153.5 |
| Q4_1 | 115.15 GiB | 117.4 | 118.2 | 119.8 | 123.0 | 129.4 | 142.1 |
| Q4_K_L | 112.07 GiB | 114.3 | 115.1 | 116.7 | 119.9 | 126.3 | 139.0 |
| Q4_K_M | 111.70 GiB | 114.0 | 114.8 | 116.4 | 119.6 | 125.9 | 138.7 |
| Q4_K | 110.56 GiB | 112.8 | 113.6 | 115.2 | 118.4 | 124.8 | 137.5 |
| Q4_K_S | 107.70 GiB | 110.0 | 110.8 | 112.4 | 115.6 | 121.9 | 134.7 |
| Q4_0 | 105.60 GiB | 107.9 | 108.7 | 110.3 | 113.5 | 119.8 | 132.6 |
| IQ4_NL | 103.97 GiB | 106.2 | 107.0 | 108.6 | 111.8 | 118.2 | 131.0 |
| IQ4_XS | 98.34 GiB | 100.6 | 101.4 | 103.0 | 106.2 | 112.6 | 125.3 |
| Q3_K_L | 87.26 GiB | 89.5 | 90.3 | 91.9 | 95.1 | 101.5 | 114.2 |
| Q3_K_M | 84.04 GiB | 86.3 | 87.1 | 88.7 | 91.9 | 98.3 | 111.0 |
| IQ3_M | 83.98 GiB | 86.2 | 87.0 | 88.6 | 91.8 | 98.2 | 111.0 |
| Q3_K_S | 80.06 GiB | 82.3 | 83.1 | 84.7 | 87.9 | 94.3 | 107.0 |
| IQ3_XS | 75.46 GiB | 77.7 | 78.5 | 80.1 | 83.3 | 89.7 | 102.4 |
| IQ3_XXS | 72.82 GiB | 75.1 | 75.9 | 77.5 | 80.7 | 87.0 | 99.8 |
| Q2_K_L | 65.26 GiB | 67.5 | 68.3 | 69.9 | 73.1 | 79.5 | 92.2 |
| Q2_K | 64.78 GiB | 67.0 | 67.8 | 69.4 | 72.6 | 79.0 | 91.8 |
| IQ2_M | 59.59 GiB | 61.9 | 62.7 | 64.2 | 67.4 | 73.8 | 86.6 |
| IQ2_S | 51.70 GiB | 54.0 | 54.8 | 56.4 | 59.6 | 65.9 | 78.7 |
| IQ2_XS | 51.40 GiB | 53.7 | 54.5 | 56.1 | 59.2 | 65.6 | 78.4 |
| IQ2_XXS | 44.74 GiB | 47.0 | 47.8 | 49.4 | 52.6 | 59.0 | 71.7 |
| IQ1_M | 40.17 GiB | 42.4 | 43.2 | 44.8 | 48.0 | 54.4 | 67.1 |
| IQ1_S | 38.47 GiB | 40.7 | 41.5 | 43.1 | 46.3 | 52.7 | 65.4 |
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 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.