Can I run Gemma-4-31B-StyleTune on a GeForce RTX 5050?
Not at these settings. No indexed quantization of Gemma-4-31B-StyleTune fits GeForce RTX 5050 at any context we compute, with q4_0 KV. The smallest shipped quantization is 7.10 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 | 59.82 GiB | 61.2 | 61.4 | 61.7 | 62.4 | 63.8 | 66.7 |
| Q8_0 | 31.79 GiB | 33.2 | 33.4 | 33.7 | 34.4 | 35.8 | 38.6 |
| Q6_K_L | 26.60 GiB | 28.0 | 28.2 | 28.5 | 29.2 | 30.6 | 33.4 |
| Q6_K | 25.97 GiB | 27.4 | 27.5 | 27.9 | 28.6 | 30.0 | 32.8 |
| I1-Q6_K | 24.55 GiB | 25.9 | 26.1 | 26.5 | 27.2 | 28.6 | 31.4 |
| Q5_K_L | 22.77 GiB | 24.2 | 24.3 | 24.7 | 25.4 | 26.8 | 29.6 |
| Q5_K_M | 21.96 GiB | 23.3 | 23.5 | 23.9 | 24.6 | 26.0 | 28.8 |
| I1-Q5_K_M | 21.25 GiB | 22.6 | 22.8 | 23.2 | 23.9 | 25.3 | 28.1 |
| Q5_K_S | 20.93 GiB | 22.3 | 22.5 | 22.8 | 23.5 | 25.0 | 27.8 |
| I1-Q5_K_S | 20.75 GiB | 22.1 | 22.3 | 22.7 | 23.4 | 24.8 | 27.6 |
| Q4_K_L | 19.96 GiB | 21.4 | 21.5 | 21.9 | 22.6 | 24.0 | 26.8 |
| Q4_1 | 19.23 GiB | 20.6 | 20.8 | 21.1 | 21.8 | 23.3 | 26.1 |
| Q4_K_M | 18.99 GiB | 20.4 | 20.6 | 20.9 | 21.6 | 23.0 | 25.8 |
| I1-Q4_1 | 18.96 GiB | 20.4 | 20.5 | 20.9 | 21.6 | 23.0 | 25.8 |
| Q5_K | 18.17 GiB | 19.6 | 19.7 | 20.1 | 20.8 | 22.2 | 25.0 |
| I1-Q4_K_M | 18.14 GiB | 19.5 | 19.7 | 20.1 | 20.8 | 22.2 | 25.0 |
| Q4_K_S | 17.68 GiB | 19.1 | 19.2 | 19.6 | 20.3 | 21.7 | 24.5 |
| Q4_0 | 17.57 GiB | 19.0 | 19.1 | 19.5 | 20.2 | 21.6 | 24.4 |
| IQ4_NL | 17.53 GiB | 18.9 | 19.1 | 19.4 | 20.1 | 21.5 | 24.4 |
| I1-Q4_K_S | 17.28 GiB | 18.7 | 18.8 | 19.2 | 19.9 | 21.3 | 24.1 |
| I1-Q4_0 | 17.22 GiB | 18.6 | 18.8 | 19.1 | 19.8 | 21.2 | 24.1 |
| IQ4_XS | 16.67 GiB | 18.1 | 18.2 | 18.6 | 19.3 | 20.7 | 23.5 |
| I1-IQ4_XS | 16.28 GiB | 17.7 | 17.8 | 18.2 | 18.9 | 20.3 | 23.1 |
| Q3_K_L | 16.22 GiB | 17.6 | 17.8 | 18.1 | 18.8 | 20.2 | 23.1 |
| I1-Q3_K_L | 16.05 GiB | 17.4 | 17.6 | 18.0 | 18.7 | 20.1 | 22.9 |
| Q3_K_M | 15.39 GiB | 16.8 | 17.0 | 17.3 | 18.0 | 19.4 | 22.2 |
| I1-Q3_K_M | 14.80 GiB | 16.2 | 16.4 | 16.7 | 17.4 | 18.8 | 21.6 |
| IQ3_M | 14.65 GiB | 16.0 | 16.2 | 16.6 | 17.3 | 18.7 | 21.5 |
| I1-IQ3_M | 14.00 GiB | 15.4 | 15.6 | 15.9 | 16.6 | 18.0 | 20.8 |
| Q3_K_S | 13.91 GiB | 15.3 | 15.5 | 15.8 | 16.5 | 17.9 | 20.7 |
| Q2_K_L | 13.48 GiB | 14.9 | 15.0 | 15.4 | 16.1 | 17.5 | 20.3 |
| IQ3_XS | 13.45 GiB | 14.8 | 15.0 | 15.4 | 16.1 | 17.5 | 20.3 |
| I1-IQ3_S | 13.38 GiB | 14.8 | 14.9 | 15.3 | 16.0 | 17.4 | 20.2 |
| I1-Q3_K_S | 13.38 GiB | 14.8 | 14.9 | 15.3 | 16.0 | 17.4 | 20.2 |
| I1-IQ3_XS | 12.74 GiB | 14.1 | 14.3 | 14.7 | 15.4 | 16.8 | 19.6 |
| IQ3_XXS | 12.65 GiB | 14.0 | 14.2 | 14.6 | 15.3 | 16.7 | 19.5 |
| IQ2_M | 12.34 GiB | 13.7 | 13.9 | 14.3 | 15.0 | 16.4 | 19.2 |
| IQ3_S | 12.22 GiB | 13.6 | 13.8 | 14.1 | 14.8 | 16.2 | 19.1 |
| Q2_K | 12.19 GiB | 13.6 | 13.8 | 14.1 | 14.8 | 16.2 | 19.0 |
| I1-IQ3_XXS | 11.81 GiB | 13.2 | 13.4 | 13.7 | 14.4 | 15.8 | 18.6 |
| I1-Q2_K | 11.53 GiB | 12.9 | 13.1 | 13.4 | 14.1 | 15.6 | 18.4 |
| IQ2_XS | 11.14 GiB | 12.5 | 12.7 | 13.1 | 13.8 | 15.2 | 18.0 |
| I1-IQ2_M | 10.73 GiB | 12.1 | 12.3 | 12.6 | 13.3 | 14.8 | 17.6 |
| I1-Q2_K_S | 10.65 GiB | 12.0 | 12.2 | 12.6 | 13.3 | 14.7 | 17.5 |
| IQ2_XXS | 10.52 GiB | 11.9 | 12.1 | 12.4 | 13.1 | 14.5 | 17.4 |
| I1-IQ2_S | 10.02 GiB | 11.4 | 11.6 | 11.9 | 12.6 | 14.0 | 16.9 |
| I1-IQ2_XS | 9.31 GiB | 10.7 | 10.9 | 11.2 | 11.9 | 13.3 | 16.1 |
| IQ2_S | 9.01 GiB | 10.4 | 10.6 | 10.9 | 11.6 | 13.0 | 15.8 |
| I1-IQ2_XXS | 8.51 GiB | 9.9 | 10.1 | 10.4 | 11.1 | 12.5 | 15.3 |
| I1-IQ1_M | 7.63 GiB | 9.0 | 9.2 | 9.5 | 10.2 | 11.6 | 14.5 |
| I1-IQ1_S | 7.10 GiB | 8.5 | 8.7 | 9.0 | 9.7 | 11.1 | 13.9 |
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.