Can I run gemma-4-12B-coder-fable5-composer2.5-v1-abliterated on a GeForce RTX 3050?
Not at these settings. No indexed quantization of gemma-4-12B-coder-fable5-composer2.5-v1-abliterated fits GeForce RTX 3050 at any context we compute, with q4_0 KV. The smallest shipped quantization is 6.77 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◐ |
|---|---|---|---|---|---|---|---|
| Q8_0 | 12.68 GiB | 13.7 | 13.8 | 13.9 | 14.2 | 14.8 | 15.9 |
| Q6_K | 10.22 GiB | 11.3 | 11.3 | 11.5 | 11.8 | 12.3 | 13.4 |
| Q5_K_M | 9.07 GiB | 10.1 | 10.2 | 10.3 | 10.6 | 11.2 | 12.3 |
| Q4_K_M | 7.98 GiB | 9.0 | 9.1 | 9.2 | 9.5 | 10.1 | 11.2 |
| IQ4_XS | 7.29 GiB | 8.3 | 8.4 | 8.5 | 8.8 | 9.4 | 10.5 |
| Q3_K_M | 6.77 GiB | 7.8 | 7.9 | 8.0 | 8.3 | 8.9 | 10.0 |
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.