Can I run gemma4-26b-fiction-bf16 on a Radeon RX 6500 XT?
Not at these settings. No indexed quantization of gemma4-26b-fiction-bf16 fits Radeon RX 6500 XT at any context we compute, with q8_0 KV. The smallest shipped quantization is 7.72 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◐ |
|---|---|---|---|---|---|---|---|
| I1-Q6_K | 21.08 GiB | 22.2 | 22.3 | 22.5 | 22.8 | 23.5 | 24.8 |
| I1-Q5_K_M | 17.82 GiB | 18.9 | 19.0 | 19.2 | 19.5 | 20.2 | 21.5 |
| I1-Q5_K_S | 16.75 GiB | 17.9 | 18.0 | 18.1 | 18.5 | 19.1 | 20.5 |
| I1-Q4_K_M | 15.64 GiB | 16.8 | 16.9 | 17.0 | 17.3 | 18.0 | 19.3 |
| I1-Q4_1 | 14.87 GiB | 16.0 | 16.1 | 16.2 | 16.6 | 17.2 | 18.6 |
| I1-Q4_K_S | 14.40 GiB | 15.5 | 15.6 | 15.8 | 16.1 | 16.8 | 18.1 |
| I1-Q4_0 | 13.49 GiB | 14.6 | 14.7 | 14.9 | 15.2 | 15.9 | 17.2 |
| I1-IQ4_XS | 12.96 GiB | 14.1 | 14.2 | 14.3 | 14.7 | 15.3 | 16.7 |
| I1-Q3_K_L | 12.88 GiB | 14.0 | 14.1 | 14.3 | 14.6 | 15.2 | 16.6 |
| I1-Q3_K_M | 12.37 GiB | 13.5 | 13.6 | 13.7 | 14.1 | 14.7 | 16.1 |
| I1-IQ3_M | 11.54 GiB | 12.7 | 12.8 | 12.9 | 13.2 | 13.9 | 15.2 |
| I1-IQ3_S | 11.38 GiB | 12.5 | 12.6 | 12.8 | 13.1 | 13.8 | 15.1 |
| I1-Q3_K_S | 11.38 GiB | 12.5 | 12.6 | 12.8 | 13.1 | 13.8 | 15.1 |
| I1-IQ3_XS | 10.84 GiB | 12.0 | 12.0 | 12.2 | 12.5 | 13.2 | 14.5 |
| I1-IQ3_XXS | 10.55 GiB | 11.7 | 11.8 | 11.9 | 12.3 | 12.9 | 14.2 |
| I1-Q2_K_S | 9.89 GiB | 11.0 | 11.1 | 11.3 | 11.6 | 12.3 | 13.6 |
| I1-Q2_K | 9.86 GiB | 11.0 | 11.1 | 11.2 | 11.6 | 12.2 | 13.6 |
| I1-IQ2_M | 9.66 GiB | 10.8 | 10.9 | 11.0 | 11.4 | 12.0 | 13.4 |
| I1-IQ2_S | 9.20 GiB | 10.3 | 10.4 | 10.6 | 10.9 | 11.6 | 12.9 |
| I1-IQ2_XS | 9.14 GiB | 10.3 | 10.4 | 10.5 | 10.8 | 11.5 | 12.8 |
| I1-IQ2_XXS | 8.66 GiB | 9.8 | 9.9 | 10.0 | 10.4 | 11.0 | 12.4 |
| I1-IQ1_M | 8.07 GiB | 9.2 | 9.3 | 9.4 | 9.8 | 10.4 | 11.8 |
| I1-IQ1_S | 7.72 GiB | 8.8 | 8.9 | 9.1 | 9.4 | 10.1 | 11.4 |
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.