Can I run Gemma-3-27B-MeditronFO on a Radeon RX 6500 XT?
Not at these settings. No indexed quantization of Gemma-3-27B-MeditronFO fits Radeon RX 6500 XT at any context we compute, with q4_0 KV. The smallest shipped quantization is 6.26 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 | 28.13 GiB | 29.4 | 29.5 | 29.6 | 30.0 | 30.7 | 32.1 |
| I1-Q6_K | 21.72 GiB | 23.0 | 23.0 | 23.2 | 23.6 | 24.3 | 25.7 |
| Q6_K | 21.72 GiB | 23.0 | 23.0 | 23.2 | 23.6 | 24.3 | 25.7 |
| I1-Q5_K_M | 18.85 GiB | 20.1 | 20.2 | 20.3 | 20.7 | 21.4 | 22.8 |
| Q5_K_M | 18.85 GiB | 20.1 | 20.2 | 20.3 | 20.7 | 21.4 | 22.8 |
| I1-Q5_K_S | 18.38 GiB | 19.6 | 19.7 | 19.9 | 20.2 | 20.9 | 22.3 |
| Q5_K_S | 18.38 GiB | 19.6 | 19.7 | 19.9 | 20.2 | 20.9 | 22.3 |
| I1-Q4_1 | 16.81 GiB | 18.0 | 18.1 | 18.3 | 18.7 | 19.4 | 20.8 |
| I1-Q4_K_M | 16.15 GiB | 17.4 | 17.5 | 17.6 | 18.0 | 18.7 | 20.1 |
| Q4_K_M | 16.15 GiB | 17.4 | 17.5 | 17.6 | 18.0 | 18.7 | 20.1 |
| I1-Q4_K_S | 15.34 GiB | 16.6 | 16.7 | 16.8 | 17.2 | 17.9 | 19.3 |
| Q4_K_S | 15.34 GiB | 16.6 | 16.7 | 16.8 | 17.2 | 17.9 | 19.3 |
| I1-Q4_0 | 15.28 GiB | 16.5 | 16.6 | 16.8 | 17.1 | 17.8 | 19.2 |
| IQ4_XS | 14.57 GiB | 15.8 | 15.9 | 16.1 | 16.4 | 17.1 | 18.5 |
| I1-IQ4_XS | 14.45 GiB | 15.7 | 15.8 | 15.9 | 16.3 | 17.0 | 18.4 |
| I1-Q3_K_L | 14.11 GiB | 15.3 | 15.4 | 15.6 | 16.0 | 16.7 | 18.1 |
| Q3_K_L | 14.11 GiB | 15.3 | 15.4 | 15.6 | 16.0 | 16.7 | 18.1 |
| I1-Q3_K_M | 13.08 GiB | 14.3 | 14.4 | 14.6 | 14.9 | 15.6 | 17.0 |
| Q3_K_M | 13.08 GiB | 14.3 | 14.4 | 14.6 | 14.9 | 15.6 | 17.0 |
| I1-IQ3_M | 12.25 GiB | 13.5 | 13.6 | 13.7 | 14.1 | 14.8 | 16.2 |
| I1-IQ3_S | 11.90 GiB | 13.1 | 13.2 | 13.4 | 13.7 | 14.4 | 15.9 |
| I1-Q3_K_S | 11.90 GiB | 13.1 | 13.2 | 13.4 | 13.7 | 14.4 | 15.9 |
| Q3_K_S | 11.90 GiB | 13.1 | 13.2 | 13.4 | 13.7 | 14.4 | 15.9 |
| I1-IQ3_XS | 11.33 GiB | 12.6 | 12.7 | 12.8 | 13.2 | 13.9 | 15.3 |
| I1-IQ3_XXS | 10.54 GiB | 11.8 | 11.9 | 12.0 | 12.4 | 13.1 | 14.5 |
| I1-Q2_K | 10.21 GiB | 11.4 | 11.5 | 11.7 | 12.1 | 12.8 | 14.2 |
| Q2_K | 10.21 GiB | 11.4 | 11.5 | 11.7 | 12.1 | 12.8 | 14.2 |
| I1-Q2_K_S | 9.52 GiB | 10.8 | 10.8 | 11.0 | 11.4 | 12.1 | 13.5 |
| I1-IQ2_M | 9.41 GiB | 10.6 | 10.7 | 10.9 | 11.3 | 12.0 | 13.4 |
| I1-IQ2_S | 8.74 GiB | 10.0 | 10.1 | 10.2 | 10.6 | 11.3 | 12.7 |
| I1-IQ2_XS | 8.29 GiB | 9.5 | 9.6 | 9.8 | 10.1 | 10.8 | 12.2 |
| I1-IQ2_XXS | 7.59 GiB | 8.8 | 8.9 | 9.1 | 9.4 | 10.1 | 11.5 |
| I1-IQ1_M | 6.76 GiB | 8.0 | 8.1 | 8.3 | 8.6 | 9.3 | 10.7 |
| I1-IQ1_S | 6.26 GiB | 7.5 | 7.6 | 7.8 | 8.1 | 8.8 | 10.2 |
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.