Can I run OlympicCoder-32B on a Radeon RX 6500 XT?
Not at these settings. No indexed quantization of OlympicCoder-32B fits Radeon RX 6500 XT at any context we compute, with q4_0 KV. The smallest shipped quantization is 8.41 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 | 32.43 GiB | 33.7 | 34.0 | 34.6 | 35.7 | 37.9 | 42.4 |
| Q6_K_L | 25.39 GiB | 26.7 | 27.0 | 27.5 | 28.6 | 30.9 | 35.4 |
| Q6_K | 25.04 GiB | 26.3 | 26.6 | 27.2 | 28.3 | 30.5 | 35.0 |
| Q5_K_L | 22.11 GiB | 23.4 | 23.7 | 24.2 | 25.4 | 27.6 | 32.1 |
| Q5_K_M | 21.66 GiB | 22.9 | 23.2 | 23.8 | 24.9 | 27.2 | 31.7 |
| Q5_K_S | 21.08 GiB | 22.4 | 22.6 | 23.2 | 24.3 | 26.6 | 31.1 |
| Q4_1 | 19.22 GiB | 20.5 | 20.8 | 21.3 | 22.5 | 24.7 | 29.2 |
| Q4_K_L | 19.03 GiB | 20.3 | 20.6 | 21.2 | 22.3 | 24.5 | 29.0 |
| Q4_K_M | 18.49 GiB | 19.8 | 20.0 | 20.6 | 21.7 | 24.0 | 28.5 |
| Q4_K_S | 17.49 GiB | 18.8 | 19.1 | 19.6 | 20.7 | 23.0 | 27.5 |
| Q4_0 | 17.43 GiB | 18.7 | 19.0 | 19.6 | 20.7 | 22.9 | 27.4 |
| IQ4_NL | 17.40 GiB | 18.7 | 19.0 | 19.5 | 20.6 | 22.9 | 27.4 |
| IQ4_XS | 16.48 GiB | 17.8 | 18.0 | 18.6 | 19.7 | 22.0 | 26.5 |
| Q3_K_L | 16.06 GiB | 17.3 | 17.6 | 18.2 | 19.3 | 21.6 | 26.1 |
| Q3_K_M | 14.84 GiB | 16.1 | 16.4 | 17.0 | 18.1 | 20.3 | 24.8 |
| IQ3_M | 13.79 GiB | 15.1 | 15.4 | 15.9 | 17.0 | 19.3 | 23.8 |
| Q3_K_S | 13.40 GiB | 14.7 | 15.0 | 15.5 | 16.7 | 18.9 | 23.4 |
| IQ3_XS | 12.76 GiB | 14.0 | 14.3 | 14.9 | 16.0 | 18.3 | 22.8 |
| Q2_K_L | 12.18 GiB | 13.5 | 13.7 | 14.3 | 15.4 | 17.7 | 22.2 |
| IQ3_XXS | 11.96 GiB | 13.2 | 13.5 | 14.1 | 15.2 | 17.5 | 22.0 |
| Q2_K | 11.47 GiB | 12.7 | 13.0 | 13.6 | 14.7 | 17.0 | 21.5 |
| IQ2_M | 10.49 GiB | 11.8 | 12.1 | 12.6 | 13.7 | 16.0 | 20.5 |
| IQ2_S | 9.67 GiB | 11.0 | 11.2 | 11.8 | 12.9 | 15.2 | 19.7 |
| IQ2_XS | 9.27 GiB | 10.6 | 10.8 | 11.4 | 12.5 | 14.8 | 19.3 |
| IQ2_XXS | 8.41 GiB | 9.7 | 10.0 | 10.5 | 11.7 | 13.9 | 18.4 |
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 131,072-token window rather than the full context.