Can I run gemma-2-27b-it on a GeForce RTX 3050?
Not at these settings. No indexed quantization of gemma-2-27b-it fits GeForce RTX 3050 at any context we compute, with q4_0 KV. The smallest shipped quantization is 5.71 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◐ |
|---|---|---|---|---|---|---|---|
| F32 | 101.43 GiB | 102.8 | 103.0 | 103.4 | 104.2 | 105.8 | 109.1 |
| BF16 | 50.72 GiB | 52.1 | 52.3 | 52.7 | 53.5 | 55.1 | 58.4 |
| Q8_0 | 26.95 GiB | 28.3 | 28.5 | 28.9 | 29.7 | 31.3 | 34.6 |
| Q6_K_L | 21.08 GiB | 22.4 | 22.6 | 23.0 | 23.9 | 25.5 | 28.7 |
| Q6_K | 20.81 GiB | 22.1 | 22.4 | 22.8 | 23.6 | 25.2 | 28.4 |
| Q5_K_L | 18.34 GiB | 19.7 | 19.9 | 20.3 | 21.1 | 22.7 | 26.0 |
| Q5_K_M | 18.08 GiB | 19.4 | 19.6 | 20.0 | 20.9 | 22.5 | 25.7 |
| Q5_K | 18.08 GiB | 19.4 | 19.6 | 20.0 | 20.9 | 22.5 | 25.7 |
| Q5_K_S | 17.59 GiB | 18.9 | 19.2 | 19.6 | 20.4 | 22.0 | 25.2 |
| Q4_K_L | 15.77 GiB | 17.1 | 17.3 | 17.7 | 18.5 | 20.2 | 23.4 |
| Q4_K_M | 15.50 GiB | 16.8 | 17.1 | 17.5 | 18.3 | 19.9 | 23.1 |
| Q4_K | 15.50 GiB | 16.8 | 17.1 | 17.5 | 18.3 | 19.9 | 23.1 |
| Q4_K_S | 14.66 GiB | 16.0 | 16.2 | 16.6 | 17.4 | 19.1 | 22.3 |
| IQ4_NL | 14.56 GiB | 15.9 | 16.1 | 16.5 | 17.3 | 19.0 | 22.2 |
| IQ4_XS | 13.80 GiB | 15.1 | 15.4 | 15.8 | 16.6 | 18.2 | 21.4 |
| Q3_K_L | 13.52 GiB | 14.9 | 15.1 | 15.5 | 16.3 | 17.9 | 21.2 |
| Q3_K | 12.50 GiB | 13.8 | 14.1 | 14.5 | 15.3 | 16.9 | 20.1 |
| Q3_K_M | 12.50 GiB | 13.8 | 14.1 | 14.5 | 15.3 | 16.9 | 20.1 |
| IQ3_M | 11.60 GiB | 12.9 | 13.2 | 13.6 | 14.4 | 16.0 | 19.2 |
| IQ3_S | 11.33 GiB | 12.7 | 12.9 | 13.3 | 14.1 | 15.7 | 19.0 |
| Q3_K_S | 11.33 GiB | 12.7 | 12.9 | 13.3 | 14.1 | 15.7 | 19.0 |
| IQ3_XS | 10.76 GiB | 12.1 | 12.3 | 12.7 | 13.5 | 15.2 | 18.4 |
| IQ3_XXS | 10.01 GiB | 11.4 | 11.6 | 12.0 | 12.8 | 14.4 | 17.6 |
| Q2_K_L | 10.00 GiB | 11.3 | 11.6 | 12.0 | 12.8 | 14.4 | 17.6 |
| Q2_K | 9.73 GiB | 11.1 | 11.3 | 11.7 | 12.5 | 14.1 | 17.4 |
| Q2_K_S | 9.06 GiB | 10.4 | 10.6 | 11.0 | 11.8 | 13.5 | 16.7 |
| IQ2_M | 8.75 GiB | 10.1 | 10.3 | 10.7 | 11.5 | 13.1 | 16.4 |
| IQ2_S | 8.06 GiB | 9.4 | 9.6 | 10.0 | 10.8 | 12.5 | 15.7 |
| IQ2_XS | 7.82 GiB | 9.2 | 9.4 | 9.8 | 10.6 | 12.2 | 15.5 |
| IQ2_XXS | 7.10 GiB | 8.4 | 8.7 | 9.1 | 9.9 | 11.5 | 14.7 |
| IQ1_M | 6.23 GiB | 7.6 | 7.8 | 8.2 | 9.0 | 10.6 | 13.9 |
| IQ1_S | 5.71 GiB | 7.0 | 7.3 | 7.7 | 8.5 | 10.1 | 13.3 |
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.