Can I run gemma-4-19b-a4b-it-REAP-heretic on a GeForce RTX 3050?
Not at these settings. No indexed quantization of gemma-4-19b-a4b-it-REAP-heretic fits GeForce RTX 3050 at any context we compute, with q4_0 KV. The smallest shipped quantization is 5.70 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◐ |
|---|---|---|---|---|---|---|---|
| Q6_K | 26.25 GiB | 27.2 | 27.2 | 27.3 | 27.5 | 27.8 | 28.5 |
| I1-Q6_K | 15.38 GiB | 16.3 | 16.3 | 16.4 | 16.6 | 17.0 | 17.7 |
| I1-Q5_K_M | 13.03 GiB | 13.9 | 14.0 | 14.1 | 14.2 | 14.6 | 15.3 |
| I1-Q5_K_S | 12.27 GiB | 13.2 | 13.2 | 13.3 | 13.5 | 13.8 | 14.5 |
| I1-Q4_K_M | 11.45 GiB | 12.4 | 12.4 | 12.5 | 12.7 | 13.0 | 13.7 |
| I1-Q4_1 | 10.91 GiB | 11.8 | 11.9 | 12.0 | 12.1 | 12.5 | 13.2 |
| I1-Q4_K_S | 10.56 GiB | 11.5 | 11.5 | 11.6 | 11.8 | 12.1 | 12.8 |
| I1-Q4_0 | 9.92 GiB | 10.8 | 10.9 | 11.0 | 11.1 | 11.5 | 12.2 |
| I1-IQ4_NL | 9.88 GiB | 10.8 | 10.8 | 10.9 | 11.1 | 11.5 | 12.2 |
| I1-IQ4_XS | 9.53 GiB | 10.4 | 10.5 | 10.6 | 10.8 | 11.1 | 11.8 |
| I1-Q3_K_L | 9.48 GiB | 10.4 | 10.4 | 10.5 | 10.7 | 11.0 | 11.8 |
| Q4_K_M | 9.38 GiB | 10.3 | 10.3 | 10.4 | 10.6 | 11.0 | 11.7 |
| I1-Q3_K_M | 9.10 GiB | 10.0 | 10.1 | 10.1 | 10.3 | 10.7 | 11.4 |
| I1-IQ3_M | 8.51 GiB | 9.4 | 9.5 | 9.6 | 9.7 | 10.1 | 10.8 |
| Q3_K_M | 8.48 GiB | 9.4 | 9.4 | 9.5 | 9.7 | 10.1 | 10.8 |
| I1-IQ3_S | 8.38 GiB | 9.3 | 9.3 | 9.4 | 9.6 | 10.0 | 10.7 |
| I1-Q3_K_S | 8.38 GiB | 9.3 | 9.3 | 9.4 | 9.6 | 10.0 | 10.7 |
| I1-IQ3_XS | 7.98 GiB | 8.9 | 8.9 | 9.0 | 9.2 | 9.6 | 10.3 |
| I1-IQ3_XXS | 7.74 GiB | 8.7 | 8.7 | 8.8 | 9.0 | 9.3 | 10.0 |
| I1-Q2_K_S | 7.29 GiB | 8.2 | 8.2 | 8.3 | 8.5 | 8.9 | 9.6 |
| I1-Q2_K | 7.28 GiB | 8.2 | 8.2 | 8.3 | 8.5 | 8.9 | 9.6 |
| I1-IQ2_M | 7.12 GiB | 8.0 | 8.1 | 8.2 | 8.3 | 8.7 | 9.4 |
| I1-IQ2_S | 6.78 GiB | 7.7 | 7.7 | 7.8 | 8.0 | 8.4 | 9.1 |
| I1-IQ2_XS | 6.73 GiB | 7.6 | 7.7 | 7.8 | 8.0 | 8.3 | 9.0 |
| I1-IQ2_XXS | 6.38 GiB | 7.3 | 7.3 | 7.4 | 7.6 | 8.0 | 8.7 |
| I1-IQ1_M | 5.96 GiB | 6.9 | 6.9 | 7.0 | 7.2 | 7.5 | 8.2 |
| I1-IQ1_S | 5.70 GiB | 6.6 | 6.7 | 6.7 | 6.9 | 7.3 | 8.0 |
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 1,024-token window rather than the full context.