Can I run Gemma-The-Writer-N-Restless-Quill-10B-Uncensored on a Radeon RX 6500 XT?
Not at these settings. No indexed quantization of Gemma-The-Writer-N-Restless-Quill-10B-Uncensored fits Radeon RX 6500 XT at any context we compute, with q8_0 KV. The smallest shipped quantization is 2.37 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 | 31.41 GiB | 33.1 | 33.5 | 34.3 | 35.8 | 38.9 | 45.0 |
| Q6_K | 25.24 GiB | 26.9 | 27.4 | 28.1 | 29.7 | 32.7 | 38.8 |
| IQ4_XS | 17.99 GiB | 19.7 | 20.1 | 20.9 | 22.4 | 25.5 | 31.6 |
| Q2_K | 13.24 GiB | 14.9 | 15.4 | 16.1 | 17.7 | 20.7 | 26.8 |
| I1-Q6_K | 7.67 GiB | 9.4 | 9.8 | 10.6 | 12.1 | 15.2 | 21.3 |
| I1-Q5_K_M | 6.71 GiB | 8.4 | 8.8 | 9.6 | 11.1 | 14.2 | 20.3 |
| Q5_K_M | 6.71 GiB | 8.4 | 8.8 | 9.6 | 11.1 | 14.2 | 20.3 |
| I1-Q5_K_S | 6.55 GiB | 8.2 | 8.7 | 9.4 | 11.0 | 14.0 | 20.1 |
| Q5_K_S | 6.55 GiB | 8.2 | 8.7 | 9.4 | 11.0 | 14.0 | 20.1 |
| I1-Q4_K_M | 5.80 GiB | 7.5 | 7.9 | 8.7 | 10.2 | 13.3 | 19.4 |
| Q4_K_M | 5.68 GiB | 7.4 | 7.8 | 8.6 | 10.1 | 13.2 | 19.3 |
| I1-Q4_K_S | 5.52 GiB | 7.2 | 7.7 | 8.4 | 9.9 | 13.0 | 19.1 |
| I1-Q4_0 | 5.50 GiB | 7.2 | 7.6 | 8.4 | 9.9 | 13.0 | 19.1 |
| Q4_K_S | 5.40 GiB | 7.1 | 7.5 | 8.3 | 9.8 | 12.9 | 19.0 |
| I1-IQ4_XS | 5.22 GiB | 6.9 | 7.4 | 8.1 | 9.6 | 12.7 | 18.8 |
| I1-Q3_K_L | 5.17 GiB | 6.9 | 7.3 | 8.1 | 9.6 | 12.6 | 18.8 |
| I1-Q3_K_M | 4.79 GiB | 6.5 | 6.9 | 7.7 | 9.2 | 12.3 | 18.4 |
| Q3_K_M | 4.57 GiB | 6.3 | 6.7 | 7.5 | 9.0 | 12.0 | 18.2 |
| I1-IQ3_M | 4.51 GiB | 6.2 | 6.6 | 7.4 | 8.9 | 12.0 | 18.1 |
| I1-IQ3_S | 4.36 GiB | 6.1 | 6.5 | 7.3 | 8.8 | 11.8 | 17.9 |
| I1-Q3_K_S | 4.36 GiB | 6.1 | 6.5 | 7.3 | 8.8 | 11.8 | 17.9 |
| I1-IQ3_XS | 4.16 GiB | 5.9 | 6.3 | 7.1 | 8.6 | 11.6 | 17.7 |
| Q3_K_S | 4.14 GiB | 5.8 | 6.3 | 7.0 | 8.6 | 11.6 | 17.7 |
| I1-IQ3_XXS | 3.82 GiB | 5.5 | 5.9 | 6.7 | 8.2 | 11.3 | 17.4 |
| I1-Q2_K | 3.81 GiB | 5.5 | 5.9 | 6.7 | 8.2 | 11.3 | 17.4 |
| I1-IQ2_M | 3.44 GiB | 5.1 | 5.6 | 6.3 | 7.9 | 10.9 | 17.0 |
| I1-IQ2_S | 3.22 GiB | 4.9 | 5.3 | 6.1 | 7.6 | 10.7 | 16.8 |
| I1-IQ2_XS | 3.07 GiB | 4.8 | 5.2 | 6.0 | 7.5 | 10.6 | 16.7 |
| I1-IQ2_XXS | 2.82 GiB | 4.5 | 5.0 | 5.7 | 7.2 | 10.3 | 16.4 |
| I1-IQ1_M | 2.54 GiB | 4.2 | 4.7 | 5.4 | 7.0 | 10.0 | 16.1 |
| I1-IQ1_S | 2.37 GiB | 4.1 | 4.5 | 5.3 | 6.8 | 9.8 | 16.0 |
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 4,096-token window rather than the full context.