Can I run solar-pro-preview-instruct on a GeForce RTX 3050?
Not at these settings. No indexed quantization of solar-pro-preview-instruct fits GeForce RTX 3050 at any context we compute, with q4_0 KV. The smallest shipped quantization is 4.46 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 | 21.91 GiB | 23.1 | 23.5 | 24.2 | 25.6 | 28.4 | 34.0 |
| Q6_K | 16.92 GiB | 18.1 | 18.5 | 19.2 | 20.6 | 23.4 | 29.0 |
| Q5_K_M | 14.59 GiB | 15.8 | 16.1 | 16.9 | 18.3 | 21.1 | 26.7 |
| Q5_K_S | 14.20 GiB | 15.4 | 15.8 | 16.5 | 17.9 | 20.7 | 26.3 |
| Q4_K_M | 12.40 GiB | 13.6 | 14.0 | 14.7 | 16.1 | 18.9 | 24.5 |
| Q4_K_S | 11.73 GiB | 12.9 | 13.3 | 14.0 | 15.4 | 18.2 | 23.8 |
| IQ4_XS | 11.06 GiB | 12.3 | 12.6 | 13.3 | 14.7 | 17.5 | 23.2 |
| Q3_K_L | 10.84 GiB | 12.0 | 12.4 | 13.1 | 14.5 | 17.3 | 22.9 |
| Q3_K_M | 9.95 GiB | 11.2 | 11.5 | 12.2 | 13.6 | 16.4 | 22.1 |
| Q3_K_S | 8.92 GiB | 10.1 | 10.5 | 11.2 | 12.6 | 15.4 | 21.0 |
| IQ3_XS | 8.50 GiB | 9.7 | 10.1 | 10.8 | 12.2 | 15.0 | 20.6 |
| Q2_K | 7.65 GiB | 8.9 | 9.2 | 9.9 | 11.3 | 14.1 | 19.8 |
| IQ2_XS | 6.16 GiB | 7.4 | 7.7 | 8.4 | 9.8 | 12.6 | 18.3 |
| IQ1_M | 4.87 GiB | 6.1 | 6.4 | 7.1 | 8.5 | 11.4 | 17.0 |
| IQ1_S | 4.46 GiB | 5.7 | 6.0 | 6.7 | 8.1 | 10.9 | 16.6 |
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.