Can I run gemma-3-12b-it on a RTX A400?
Not at these settings. No indexed quantization of gemma-3-12b-it fits RTX A400 at any context we compute, with q8_0 KV. The smallest shipped quantization is 2.85 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 | 21.92 GiB | 23.4 | 23.5 | 23.7 | 24.3 | 25.3 | 27.5 |
| Q8_0 | 11.65 GiB | 13.1 | 13.2 | 13.5 | 14.0 | 15.1 | 17.2 |
| Q6_K | 9.00 GiB | 10.4 | 10.6 | 10.8 | 11.4 | 12.4 | 14.5 |
| Q5_K_M | 7.87 GiB | 9.3 | 9.4 | 9.7 | 10.2 | 11.3 | 13.4 |
| Q5_K_S | 7.67 GiB | 9.1 | 9.2 | 9.5 | 10.0 | 11.1 | 13.2 |
| Q4_0 | 7.52 GiB | 8.9 | 9.1 | 9.3 | 9.9 | 10.9 | 13.1 |
| Q4_1 | 7.04 GiB | 8.5 | 8.6 | 8.9 | 9.4 | 10.5 | 12.6 |
| Q4_K_M | 6.80 GiB | 8.2 | 8.4 | 8.6 | 9.2 | 10.2 | 12.3 |
| Q4_K_S | 6.46 GiB | 7.9 | 8.0 | 8.3 | 8.8 | 9.9 | 12.0 |
| IQ4_NL | 6.41 GiB | 7.8 | 8.0 | 8.2 | 8.8 | 9.8 | 12.0 |
| IQ4_XS | 6.10 GiB | 7.5 | 7.7 | 7.9 | 8.5 | 9.5 | 11.6 |
| Q3_K_L | 6.03 GiB | 7.5 | 7.6 | 7.9 | 8.4 | 9.5 | 11.6 |
| Q3_K_M | 5.60 GiB | 7.0 | 7.2 | 7.4 | 8.0 | 9.0 | 11.1 |
| Q3_K_S | 5.08 GiB | 6.5 | 6.6 | 6.9 | 7.4 | 8.5 | 10.6 |
| UD-IQ3_XXS | 4.50 GiB | 5.9 | 6.1 | 6.3 | 6.9 | 7.9 | 10.0 |
| Q2_K | 4.44 GiB | 5.9 | 6.0 | 6.3 | 6.8 | 7.9 | 10.0 |
| Q2_K_L | 4.44 GiB | 5.9 | 6.0 | 6.3 | 6.8 | 7.9 | 10.0 |
| UD-IQ2_M | 4.07 GiB | 5.5 | 5.6 | 5.9 | 6.4 | 7.5 | 9.6 |
| UD-IQ2_XXS | 3.36 GiB | 4.8 | 4.9 | 5.2 | 5.7 | 6.8 | 8.9 |
| UD-IQ1_M | 3.03 GiB | 4.5 | 4.6 | 4.9 | 5.4 | 6.4 | 8.6 |
| UD-IQ1_S | 2.85 GiB | 4.3 | 4.4 | 4.7 | 5.2 | 6.3 | 8.4 |
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.