Can I run gemma-4-12B on a Radeon RX 6500 XT?
Not at these settings. No indexed quantization of gemma-4-12B fits Radeon RX 6500 XT at any context we compute, with q4_0 KV. The smallest shipped quantization is 2.72 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◐ |
|---|---|---|---|---|---|---|---|
| BF16 | 22.20 GiB | 23.3 | 23.4 | 23.6 | 23.8 | 24.4 | 25.5 |
| Q8_0 | 11.80 GiB | 12.9 | 13.0 | 13.2 | 13.4 | 14.0 | 15.1 |
| Q6_K_M | 9.34 GiB | 10.5 | 10.6 | 10.7 | 11.0 | 11.5 | 12.7 |
| Q6_K | 9.11 GiB | 10.3 | 10.3 | 10.5 | 10.8 | 11.3 | 12.4 |
| Q5_1 | 8.63 GiB | 9.8 | 9.8 | 10.0 | 10.3 | 10.8 | 12.0 |
| Q5_K_M | 8.37 GiB | 9.5 | 9.6 | 9.7 | 10.0 | 10.6 | 11.7 |
| Q5_0 | 7.99 GiB | 9.1 | 9.2 | 9.4 | 9.6 | 10.2 | 11.3 |
| Q4_K_M | 7.21 GiB | 8.4 | 8.4 | 8.6 | 8.8 | 9.4 | 10.5 |
| Q4_1 | 7.01 GiB | 8.2 | 8.2 | 8.4 | 8.6 | 9.2 | 10.3 |
| Q4_K_S | 6.78 GiB | 7.9 | 8.0 | 8.1 | 8.4 | 9.0 | 10.1 |
| Q4_0 | 6.72 GiB | 7.9 | 7.9 | 8.1 | 8.4 | 8.9 | 10.1 |
| IQ4_NL | 6.26 GiB | 7.4 | 7.5 | 7.6 | 7.9 | 8.5 | 9.6 |
| IQ4_XS | 6.18 GiB | 7.3 | 7.4 | 7.5 | 7.8 | 8.4 | 9.5 |
| Q3_K_M | 5.87 GiB | 7.0 | 7.1 | 7.2 | 7.5 | 8.1 | 9.2 |
| Q3_K_S | 5.66 GiB | 6.8 | 6.9 | 7.0 | 7.3 | 7.9 | 9.0 |
| IQ3_M | 5.41 GiB | 6.6 | 6.6 | 6.8 | 7.1 | 7.6 | 8.7 |
| IQ3_XS | 4.91 GiB | 6.1 | 6.1 | 6.3 | 6.6 | 7.1 | 8.2 |
| IQ3_XXS | 4.67 GiB | 5.8 | 5.9 | 6.0 | 6.3 | 6.9 | 8.0 |
| Q2_K_M | 4.57 GiB | 5.7 | 5.8 | 5.9 | 6.2 | 6.8 | 7.9 |
| Q2_K_S | 4.38 GiB | 5.5 | 5.6 | 5.7 | 6.0 | 6.6 | 7.7 |
| IQ2_M | 4.16 GiB | 5.3 | 5.4 | 5.5 | 5.8 | 6.4 | 7.5 |
| IQ2_S | 3.93 GiB | 5.1 | 5.2 | 5.3 | 5.6 | 6.1 | 7.3 |
| IQ2_XS | 3.68 GiB | 4.8 | 4.9 | 5.0 | 5.3 | 5.9 | 7.0 |
| IQ2_XXS | 3.38 GiB | 4.5 | 4.6 | 4.7 | 5.0 | 5.6 | 6.7 |
| IQ1_M | 3.14 GiB | 4.3 | 4.4 | 4.5 | 4.8 | 5.3 | 6.5 |
| IQ1_S | 2.72 GiB | 3.9 | 3.9 | 4.1 | 4.4 | 4.9 | 6.1 |
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.