Can I run Gemma-4-AssGuard-31B on a Radeon RX 6500 XT?
Not at these settings. No indexed quantization of Gemma-4-AssGuard-31B fits Radeon RX 6500 XT at any context we compute, with q8_0 KV. The smallest shipped quantization is 6.67 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 | 30.39 GiB | 32.3 | 32.7 | 33.3 | 34.7 | 37.3 | 42.6 |
| I1-Q6_K | 23.47 GiB | 25.4 | 25.7 | 26.4 | 27.7 | 30.4 | 35.7 |
| Q6_K | 23.47 GiB | 25.4 | 25.7 | 26.4 | 27.7 | 30.4 | 35.7 |
| I1-Q5_K_M | 20.35 GiB | 22.3 | 22.6 | 23.3 | 24.6 | 27.3 | 32.6 |
| Q5_K_M | 20.35 GiB | 22.3 | 22.6 | 23.3 | 24.6 | 27.3 | 32.6 |
| I1-Q5_K_S | 19.85 GiB | 21.8 | 22.1 | 22.8 | 24.1 | 26.8 | 32.1 |
| Q5_K_S | 19.85 GiB | 21.8 | 22.1 | 22.8 | 24.1 | 26.8 | 32.1 |
| I1-Q4_1 | 18.14 GiB | 20.1 | 20.4 | 21.1 | 22.4 | 25.1 | 30.4 |
| I1-Q4_K_M | 17.40 GiB | 19.3 | 19.7 | 20.3 | 21.7 | 24.3 | 29.6 |
| Q4_K_M | 17.40 GiB | 19.3 | 19.7 | 20.3 | 21.7 | 24.3 | 29.6 |
| I1-Q4_K_S | 16.54 GiB | 18.5 | 18.8 | 19.5 | 20.8 | 23.5 | 28.8 |
| Q4_K_S | 16.54 GiB | 18.5 | 18.8 | 19.5 | 20.8 | 23.5 | 28.8 |
| I1-Q4_0 | 16.49 GiB | 18.4 | 18.8 | 19.4 | 20.7 | 23.4 | 28.7 |
| IQ4_XS | 15.70 GiB | 17.6 | 18.0 | 18.6 | 20.0 | 22.6 | 27.9 |
| I1-IQ4_XS | 15.59 GiB | 17.5 | 17.9 | 18.5 | 19.8 | 22.5 | 27.8 |
| I1-Q3_K_L | 15.49 GiB | 17.4 | 17.8 | 18.4 | 19.7 | 22.4 | 27.7 |
| Q3_K_L | 15.49 GiB | 17.4 | 17.8 | 18.4 | 19.7 | 22.4 | 27.7 |
| I1-Q3_K_M | 14.24 GiB | 16.2 | 16.5 | 17.2 | 18.5 | 21.2 | 26.5 |
| Q3_K_M | 14.24 GiB | 16.2 | 16.5 | 17.2 | 18.5 | 21.2 | 26.5 |
| I1-IQ3_M | 13.43 GiB | 15.4 | 15.7 | 16.4 | 17.7 | 20.3 | 25.7 |
| I1-IQ3_S | 12.82 GiB | 14.8 | 15.1 | 15.7 | 17.1 | 19.7 | 25.0 |
| I1-Q3_K_S | 12.82 GiB | 14.8 | 15.1 | 15.7 | 17.1 | 19.7 | 25.0 |
| Q3_K_S | 12.82 GiB | 14.8 | 15.1 | 15.7 | 17.1 | 19.7 | 25.0 |
| I1-IQ3_XS | 12.17 GiB | 14.1 | 14.4 | 15.1 | 16.4 | 19.1 | 24.4 |
| I1-IQ3_XXS | 11.25 GiB | 13.2 | 13.5 | 14.2 | 15.5 | 18.2 | 23.5 |
| I1-Q2_K | 11.10 GiB | 13.0 | 13.4 | 14.0 | 15.4 | 18.0 | 23.3 |
| Q2_K | 11.10 GiB | 13.0 | 13.4 | 14.0 | 15.4 | 18.0 | 23.3 |
| I1-Q2_K_S | 10.22 GiB | 12.2 | 12.5 | 13.2 | 14.5 | 17.1 | 22.5 |
| I1-IQ2_M | 10.17 GiB | 12.1 | 12.4 | 13.1 | 14.4 | 17.1 | 22.4 |
| I1-IQ2_S | 9.46 GiB | 11.4 | 11.7 | 12.4 | 13.7 | 16.4 | 21.7 |
| I1-IQ2_XS | 8.88 GiB | 10.8 | 11.1 | 11.8 | 13.1 | 15.8 | 21.1 |
| I1-IQ2_XXS | 8.08 GiB | 10.0 | 10.3 | 11.0 | 12.3 | 15.0 | 20.3 |
| I1-IQ1_M | 7.20 GiB | 9.1 | 9.5 | 10.1 | 11.5 | 14.1 | 19.4 |
| I1-IQ1_S | 6.67 GiB | 8.6 | 8.9 | 9.6 | 10.9 | 13.6 | 18.9 |
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.