Can I run gemma-4-E4B-it-qat-q4_0-unquantized-heretic on a Radeon RX 6500 XT?
Not at these settings. No indexed quantization of gemma-4-E4B-it-qat-q4_0-unquantized-heretic fits Radeon RX 6500 XT at any context we compute, with q4_0 KV. The smallest shipped quantization is 3.06 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 | 5.75 GiB | 6.7 | 6.7 | 6.7 | 6.8 | 6.9 | 7.2 |
| I1-Q5_K_M | 5.33 GiB | 6.3 | 6.3 | 6.3 | 6.4 | 6.5 | 6.8 |
| I1-Q5_K_S | 5.26 GiB | 6.2 | 6.2 | 6.3 | 6.3 | 6.4 | 6.7 |
| I1-Q4_1 | 5.03 GiB | 6.0 | 6.0 | 6.0 | 6.1 | 6.2 | 6.5 |
| I1-Q4_K_M | 4.94 GiB | 5.9 | 5.9 | 5.9 | 6.0 | 6.1 | 6.4 |
| I1-Q4_K_S | 4.82 GiB | 5.8 | 5.8 | 5.8 | 5.9 | 6.0 | 6.2 |
| I1-Q4_0 | 4.81 GiB | 5.8 | 5.8 | 5.8 | 5.9 | 6.0 | 6.2 |
| I1-IQ4_NL | 4.81 GiB | 5.8 | 5.8 | 5.8 | 5.9 | 6.0 | 6.2 |
| I1-IQ4_XS | 4.69 GiB | 5.6 | 5.7 | 5.7 | 5.8 | 5.9 | 6.1 |
| I1-Q3_K_L | 4.65 GiB | 5.6 | 5.6 | 5.6 | 5.7 | 5.8 | 6.1 |
| I1-Q3_K_M | 4.49 GiB | 5.4 | 5.5 | 5.5 | 5.6 | 5.7 | 5.9 |
| I1-IQ3_M | 4.37 GiB | 5.3 | 5.3 | 5.4 | 5.4 | 5.5 | 5.8 |
| I1-IQ3_S | 4.32 GiB | 5.3 | 5.3 | 5.3 | 5.4 | 5.5 | 5.7 |
| I1-Q3_K_S | 4.31 GiB | 5.3 | 5.3 | 5.3 | 5.4 | 5.5 | 5.7 |
| I1-IQ3_XS | 4.23 GiB | 5.2 | 5.2 | 5.2 | 5.3 | 5.4 | 5.7 |
| I1-Q2_K | 4.08 GiB | 5.0 | 5.0 | 5.1 | 5.1 | 5.3 | 5.5 |
| I1-Q2_K_S | 3.98 GiB | 4.9 | 4.9 | 5.0 | 5.0 | 5.2 | 5.4 |
| I1-IQ3_XXS | 3.71 GiB | 4.7 | 4.7 | 4.7 | 4.8 | 4.9 | 5.1 |
| I1-IQ2_M | 3.53 GiB | 4.5 | 4.5 | 4.5 | 4.6 | 4.7 | 5.0 |
| I1-IQ2_S | 3.42 GiB | 4.4 | 4.4 | 4.4 | 4.5 | 4.6 | 4.9 |
| I1-IQ2_XS | 3.38 GiB | 4.3 | 4.3 | 4.4 | 4.4 | 4.6 | 4.8 |
| I1-IQ2_XXS | 3.27 GiB | 4.2 | 4.2 | 4.3 | 4.3 | 4.5 | 4.7 |
| I1-IQ1_M | 3.14 GiB | 4.1 | 4.1 | 4.1 | 4.2 | 4.3 | 4.6 |
| I1-IQ1_S | 3.06 GiB | 4.0 | 4.0 | 4.1 | 4.1 | 4.2 | 4.5 |
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 512-token window rather than the full context.