Can I run solar-pro-preview-instruct on a Radeon RX 6500 XT?
Not at these settings. No indexed quantization of solar-pro-preview-instruct fits Radeon RX 6500 XT at any context we compute, with q8_0 KV. The smallest shipped quantization is 4.46 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 | 21.91 GiB | 23.5 | 24.2 | 25.5 | 28.2 | 33.5 | 44.1 |
| Q6_K | 16.92 GiB | 18.5 | 19.2 | 20.5 | 23.2 | 28.5 | 39.1 |
| Q5_K_M | 14.59 GiB | 16.2 | 16.9 | 18.2 | 20.9 | 26.2 | 36.8 |
| Q5_K_S | 14.20 GiB | 15.8 | 16.5 | 17.8 | 20.5 | 25.8 | 36.4 |
| Q4_K_M | 12.40 GiB | 14.0 | 14.7 | 16.0 | 18.7 | 24.0 | 34.6 |
| Q4_K_S | 11.73 GiB | 13.4 | 14.0 | 15.3 | 18.0 | 23.3 | 33.9 |
| IQ4_XS | 11.06 GiB | 12.7 | 13.3 | 14.7 | 17.3 | 22.6 | 33.3 |
| Q3_K_L | 10.84 GiB | 12.5 | 13.1 | 14.5 | 17.1 | 22.4 | 33.0 |
| Q3_K_M | 9.95 GiB | 11.6 | 12.2 | 13.6 | 16.2 | 21.5 | 32.2 |
| Q3_K_S | 8.92 GiB | 10.5 | 11.2 | 12.5 | 15.2 | 20.5 | 31.1 |
| IQ3_XS | 8.50 GiB | 10.1 | 10.8 | 12.1 | 14.8 | 20.1 | 30.7 |
| Q2_K | 7.65 GiB | 9.3 | 9.9 | 11.3 | 13.9 | 19.2 | 29.9 |
| IQ2_XS | 6.16 GiB | 7.8 | 8.5 | 9.8 | 12.4 | 17.7 | 28.4 |
| IQ1_M | 4.87 GiB | 6.5 | 7.2 | 8.5 | 11.1 | 16.5 | 27.1 |
| IQ1_S | 4.46 GiB | 6.1 | 6.7 | 8.1 | 10.7 | 16.0 | 26.7 |
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 2,047-token window rather than the full context.