Can I run OlympicCoder-32B on a GeForce RTX 3050?
Not at these settings. No indexed quantization of OlympicCoder-32B fits GeForce RTX 3050 at any context we compute, with q8_0 KV. The smallest shipped quantization is 8.41 GiB in weights alone, against 5.58 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 | 33.9 | 34.4 | 35.5 | 37.6 | 41.8 | 50.3 |
| Q6_K_L | 25.39 GiB | 26.8 | 27.4 | 28.4 | 30.5 | 34.8 | 43.3 |
| Q6_K | 25.04 GiB | 26.5 | 27.0 | 28.1 | 30.2 | 34.4 | 42.9 |
| Q5_K_L | 22.11 GiB | 23.5 | 24.1 | 25.1 | 27.3 | 31.5 | 40.0 |
| Q5_K_M | 21.66 GiB | 23.1 | 23.6 | 24.7 | 26.8 | 31.1 | 39.6 |
| Q5_K_S | 21.08 GiB | 22.5 | 23.0 | 24.1 | 26.2 | 30.5 | 39.0 |
| Q4_1 | 19.22 GiB | 20.7 | 21.2 | 22.2 | 24.4 | 28.6 | 37.1 |
| Q4_K_L | 19.03 GiB | 20.5 | 21.0 | 22.1 | 24.2 | 28.4 | 36.9 |
| Q4_K_M | 18.49 GiB | 19.9 | 20.4 | 21.5 | 23.6 | 27.9 | 36.4 |
| Q4_K_S | 17.49 GiB | 18.9 | 19.5 | 20.5 | 22.6 | 26.9 | 35.4 |
| Q4_0 | 17.43 GiB | 18.9 | 19.4 | 20.5 | 22.6 | 26.8 | 35.3 |
| IQ4_NL | 17.40 GiB | 18.8 | 19.4 | 20.4 | 22.5 | 26.8 | 35.3 |
| IQ4_XS | 16.48 GiB | 17.9 | 18.4 | 19.5 | 21.6 | 25.9 | 34.4 |
| Q3_K_L | 16.06 GiB | 17.5 | 18.0 | 19.1 | 21.2 | 25.5 | 34.0 |
| Q3_K_M | 14.84 GiB | 16.3 | 16.8 | 17.9 | 20.0 | 24.2 | 32.7 |
| IQ3_M | 13.79 GiB | 15.2 | 15.8 | 16.8 | 18.9 | 23.2 | 31.7 |
| Q3_K_S | 13.40 GiB | 14.8 | 15.4 | 16.4 | 18.6 | 22.8 | 31.3 |
| IQ3_XS | 12.76 GiB | 14.2 | 14.7 | 15.8 | 17.9 | 22.2 | 30.7 |
| Q2_K_L | 12.18 GiB | 13.6 | 14.1 | 15.2 | 17.3 | 21.6 | 30.1 |
| IQ3_XXS | 11.96 GiB | 13.4 | 13.9 | 15.0 | 17.1 | 21.4 | 29.9 |
| Q2_K | 11.47 GiB | 12.9 | 13.4 | 14.5 | 16.6 | 20.9 | 29.4 |
| IQ2_M | 10.49 GiB | 11.9 | 12.5 | 13.5 | 15.6 | 19.9 | 28.4 |
| IQ2_S | 9.67 GiB | 11.1 | 11.6 | 12.7 | 14.8 | 19.1 | 27.6 |
| IQ2_XS | 9.27 GiB | 10.7 | 11.2 | 12.3 | 14.4 | 18.7 | 27.2 |
| IQ2_XXS | 8.41 GiB | 9.8 | 10.4 | 11.4 | 13.6 | 17.8 | 26.3 |
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.