Can I run gemma-3-27b-it on a RTX A2000?
Not at these settings. No indexed quantization of gemma-3-27b-it fits RTX A2000 at any context we compute, with q8_0 KV. The smallest shipped quantization is 6.06 GiB in weights alone, against 5.58 GiB usable. CPU offload can still run it, slowly.
Every quantization at every context
| Quant | Weights● | 4K◐ | 8K◐ | 16K◐ | 32K◐ | 64K◐ | 128K◐ |
|---|---|---|---|---|---|---|---|
| BF16 | 50.32 GiB | 51.9 | 52.1 | 52.4 | 53.1 | 54.4 | 57.0 |
| F16 | 50.32 GiB | 51.9 | 52.1 | 52.4 | 53.0 | 54.4 | 57.0 |
| Q8_0 | 26.74 GiB | 28.3 | 28.5 | 28.8 | 29.5 | 30.8 | 33.5 |
| Q6_K_L | 20.96 GiB | 22.5 | 22.7 | 23.0 | 23.7 | 25.0 | 27.7 |
| Q6_K | 20.64 GiB | 22.2 | 22.4 | 22.7 | 23.4 | 24.7 | 27.4 |
| Q5_K_L | 18.27 GiB | 19.8 | 20.0 | 20.3 | 21.0 | 22.3 | 25.0 |
| Q5_K_M | 17.95 GiB | 19.5 | 19.7 | 20.0 | 20.7 | 22.0 | 24.7 |
| Q5_K_S | 17.48 GiB | 19.0 | 19.2 | 19.5 | 20.2 | 21.5 | 24.2 |
| Q4_1 | 15.99 GiB | 17.6 | 17.7 | 18.1 | 18.7 | 20.0 | 22.7 |
| Q4_K_L | 15.73 GiB | 17.3 | 17.5 | 17.8 | 18.5 | 19.8 | 22.4 |
| Q4_K_M | 15.41 GiB | 17.0 | 17.1 | 17.5 | 18.1 | 19.5 | 22.1 |
| Q4_K_S | 14.60 GiB | 16.2 | 16.3 | 16.7 | 17.3 | 18.7 | 21.3 |
| Q4_0 | 14.55 GiB | 16.1 | 16.3 | 16.6 | 17.3 | 18.6 | 21.3 |
| IQ4_NL | 14.50 GiB | 16.1 | 16.2 | 16.6 | 17.2 | 18.6 | 21.2 |
| IQ4_XS | 13.75 GiB | 15.3 | 15.5 | 15.8 | 16.5 | 17.8 | 20.5 |
| Q3_K_L | 13.54 GiB | 15.1 | 15.3 | 15.6 | 16.3 | 17.6 | 20.3 |
| Q3_K_M | 12.51 GiB | 14.1 | 14.3 | 14.6 | 15.2 | 16.6 | 19.2 |
| IQ3_M | 11.69 GiB | 13.3 | 13.4 | 13.8 | 14.4 | 15.7 | 18.4 |
| Q3_K_S | 11.33 GiB | 12.9 | 13.1 | 13.4 | 14.1 | 15.4 | 18.0 |
| IQ3_XS | 10.77 GiB | 12.3 | 12.5 | 12.8 | 13.5 | 14.8 | 17.5 |
| Q2_K_L | 10.10 GiB | 11.7 | 11.8 | 12.2 | 12.8 | 14.2 | 16.8 |
| UD-IQ3_XXS | 10.07 GiB | 11.6 | 11.8 | 12.1 | 12.8 | 14.1 | 16.8 |
| IQ3_XXS | 9.98 GiB | 11.6 | 11.7 | 12.0 | 12.7 | 14.0 | 16.7 |
| Q2_K | 9.78 GiB | 11.4 | 11.5 | 11.9 | 12.5 | 13.8 | 16.5 |
| UD-IQ2_M | 8.96 GiB | 10.5 | 10.7 | 11.0 | 11.7 | 13.0 | 15.7 |
| IQ2_M | 8.84 GiB | 10.4 | 10.6 | 10.9 | 11.6 | 12.9 | 15.6 |
| IQ2_S | 8.18 GiB | 9.7 | 9.9 | 10.2 | 10.9 | 12.2 | 14.9 |
| IQ2_XS | 7.86 GiB | 9.4 | 9.6 | 9.9 | 10.6 | 11.9 | 14.6 |
| UD-IQ2_XXS | 7.31 GiB | 8.9 | 9.0 | 9.4 | 10.0 | 11.4 | 14.0 |
| UD-IQ1_M | 6.51 GiB | 8.1 | 8.2 | 8.6 | 9.2 | 10.6 | 13.2 |
| UD-IQ1_S | 6.06 GiB | 7.6 | 7.8 | 8.1 | 8.8 | 10.1 | 12.8 |
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.