Can I run Pantheon-Reasoning-26B-A4B-1.1 on a GeForce RTX 5050?
Not at these settings. No indexed quantization of Pantheon-Reasoning-26B-A4B-1.1 fits GeForce RTX 5050 at any context we compute, with q8_0 KV. The smallest shipped quantization is 7.95 GiB in weights alone, against 7.44 GiB usable. CPU offload can still run it, slowly.
Every quantization at every context
| Quant | Weights● | 4K◐ | 8K◐ | 16K◐ | 32K◐ | 64K◐ | 128K◐ |
|---|---|---|---|---|---|---|---|
| BF16 | 48.41 GiB | 49.4 | 49.5 | 49.7 | 50.0 | 50.7 | 52.0 |
| Q8_0 | 25.75 GiB | 26.8 | 26.9 | 27.0 | 27.4 | 28.0 | 29.3 |
| Q6_K_L | 22.19 GiB | 23.2 | 23.3 | 23.5 | 23.8 | 24.5 | 25.8 |
| Q6_K | 21.86 GiB | 22.9 | 23.0 | 23.1 | 23.5 | 24.1 | 25.5 |
| I1-Q6_K | 21.65 GiB | 22.7 | 22.8 | 22.9 | 23.3 | 23.9 | 25.2 |
| Q5_K_L | 18.89 GiB | 19.9 | 20.0 | 20.2 | 20.5 | 21.2 | 22.5 |
| Q5_K_M | 18.47 GiB | 19.5 | 19.6 | 19.7 | 20.1 | 20.7 | 22.1 |
| I1-Q5_K_M | 18.29 GiB | 19.3 | 19.4 | 19.6 | 19.9 | 20.6 | 21.9 |
| Q5_K_S | 17.36 GiB | 18.4 | 18.5 | 18.6 | 19.0 | 19.6 | 21.0 |
| I1-Q5_K_S | 17.22 GiB | 18.3 | 18.3 | 18.5 | 18.8 | 19.5 | 20.8 |
| Q4_K_L | 16.76 GiB | 17.8 | 17.9 | 18.0 | 18.4 | 19.0 | 20.4 |
| Q4_K_M | 16.25 GiB | 17.3 | 17.4 | 17.5 | 17.9 | 18.5 | 19.9 |
| I1-Q4_K_M | 16.03 GiB | 17.1 | 17.1 | 17.3 | 17.6 | 18.3 | 19.6 |
| Q4_1 | 15.47 GiB | 16.5 | 16.6 | 16.7 | 17.1 | 17.7 | 19.1 |
| I1-Q4_1 | 15.30 GiB | 16.3 | 16.4 | 16.6 | 16.9 | 17.6 | 18.9 |
| Q4_K_S | 15.14 GiB | 16.2 | 16.3 | 16.4 | 16.7 | 17.4 | 18.7 |
| I1-Q4_K_S | 14.79 GiB | 15.8 | 15.9 | 16.1 | 16.4 | 17.1 | 18.4 |
| Q4_0 | 14.13 GiB | 15.2 | 15.2 | 15.4 | 15.7 | 16.4 | 17.7 |
| IQ4_NL | 14.08 GiB | 15.1 | 15.2 | 15.4 | 15.7 | 16.3 | 17.7 |
| I1-Q4_0 | 13.88 GiB | 14.9 | 15.0 | 15.2 | 15.5 | 16.2 | 17.5 |
| IQ4_XS | 13.59 GiB | 14.6 | 14.7 | 14.9 | 15.2 | 15.9 | 17.2 |
| I1-IQ4_XS | 13.33 GiB | 14.4 | 14.4 | 14.6 | 14.9 | 15.6 | 16.9 |
| I1-Q3_K_L | 13.17 GiB | 14.2 | 14.3 | 14.4 | 14.8 | 15.4 | 16.8 |
| I1-Q3_K_M | 12.67 GiB | 13.7 | 13.8 | 13.9 | 14.3 | 14.9 | 16.3 |
| IQ3_M | 12.66 GiB | 13.7 | 13.8 | 13.9 | 14.3 | 14.9 | 16.3 |
| Q3_K_L | 12.59 GiB | 13.6 | 13.7 | 13.9 | 14.2 | 14.9 | 16.2 |
| Q3_K_M | 12.42 GiB | 13.4 | 13.5 | 13.7 | 14.0 | 14.7 | 16.0 |
| Q3_K_S | 11.99 GiB | 13.0 | 13.1 | 13.3 | 13.6 | 14.3 | 15.6 |
| IQ3_XS | 11.88 GiB | 12.9 | 13.0 | 13.2 | 13.5 | 14.1 | 15.5 |
| I1-IQ3_M | 11.84 GiB | 12.9 | 12.9 | 13.1 | 13.4 | 14.1 | 15.4 |
| I1-IQ3_S | 11.68 GiB | 12.7 | 12.8 | 13.0 | 13.3 | 13.9 | 15.3 |
| I1-Q3_K_S | 11.68 GiB | 12.7 | 12.8 | 13.0 | 13.3 | 13.9 | 15.3 |
| IQ3_XXS | 11.62 GiB | 12.6 | 12.7 | 12.9 | 13.2 | 13.9 | 15.2 |
| I1-IQ3_XS | 11.13 GiB | 12.2 | 12.2 | 12.4 | 12.7 | 13.4 | 14.7 |
| Q2_K_L | 11.10 GiB | 12.1 | 12.2 | 12.4 | 12.7 | 13.4 | 14.7 |
| I1-IQ3_XXS | 10.84 GiB | 11.9 | 12.0 | 12.1 | 12.5 | 13.1 | 14.4 |
| Q2_K | 10.43 GiB | 11.5 | 11.5 | 11.7 | 12.0 | 12.7 | 14.0 |
| IQ2_M | 10.26 GiB | 11.3 | 11.4 | 11.5 | 11.9 | 12.5 | 13.9 |
| I1-Q2_K_S | 10.12 GiB | 11.1 | 11.2 | 11.4 | 11.7 | 12.4 | 13.7 |
| I1-Q2_K | 10.08 GiB | 11.1 | 11.2 | 11.4 | 11.7 | 12.4 | 13.7 |
| I1-IQ2_M | 9.96 GiB | 11.0 | 11.1 | 11.2 | 11.6 | 12.2 | 13.6 |
| IQ2_S | 9.82 GiB | 10.8 | 10.9 | 11.1 | 11.4 | 12.1 | 13.4 |
| IQ2_XS | 9.66 GiB | 10.7 | 10.8 | 10.9 | 11.3 | 11.9 | 13.3 |
| I1-IQ2_S | 9.49 GiB | 10.5 | 10.6 | 10.8 | 11.1 | 11.8 | 13.1 |
| I1-IQ2_XS | 9.37 GiB | 10.4 | 10.5 | 10.6 | 11.0 | 11.6 | 13.0 |
| I1-IQ2_XXS | 8.89 GiB | 9.9 | 10.0 | 10.2 | 10.5 | 11.2 | 12.5 |
| I1-IQ1_M | 8.30 GiB | 9.3 | 9.4 | 9.6 | 9.9 | 10.6 | 11.9 |
| I1-IQ1_S | 7.95 GiB | 9.0 | 9.1 | 9.2 | 9.6 | 10.2 | 11.5 |
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 1,024-token window rather than the full context.