Can I run Qwen2.5-14B-Instruct on a GeForce RTX 3050?
Not at these settings. No indexed quantization of Qwen2.5-14B-Instruct fits GeForce RTX 3050 at any context we compute, with q4_0 KV. The smallest shipped quantization is 4.99 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◐ |
|---|---|---|---|---|---|---|---|
| F16 | 27.52 GiB | 28.6 | 28.8 | 29.2 | 30.1 | 31.7 | 35.1 |
| Q8_0 | 14.62 GiB | 15.7 | 15.9 | 16.3 | 17.2 | 18.8 | 22.2 |
| Q6_K_L | 11.64 GiB | 12.7 | 12.9 | 13.3 | 14.2 | 15.9 | 19.2 |
| Q6_K | 11.29 GiB | 12.3 | 12.6 | 13.0 | 13.8 | 15.5 | 18.9 |
| Q5_K_L | 10.23 GiB | 11.3 | 11.5 | 11.9 | 12.8 | 14.5 | 17.8 |
| Q5_K_M | 9.79 GiB | 10.8 | 11.1 | 11.5 | 12.3 | 14.0 | 17.4 |
| Q5_K_S | 9.56 GiB | 10.6 | 10.8 | 11.3 | 12.1 | 13.8 | 17.2 |
| Q5_0 | 9.56 GiB | 10.6 | 10.8 | 11.3 | 12.1 | 13.8 | 17.2 |
| Q4_K_L | 8.91 GiB | 10.0 | 10.2 | 10.6 | 11.4 | 13.1 | 16.5 |
| Q4_K_M | 8.37 GiB | 9.4 | 9.6 | 10.1 | 10.9 | 12.6 | 16.0 |
| Q4_K_S | 7.98 GiB | 9.0 | 9.3 | 9.7 | 10.5 | 12.2 | 15.6 |
| Q4_0 | 7.96 GiB | 9.0 | 9.2 | 9.6 | 10.5 | 12.2 | 15.6 |
| IQ4_XS | 7.56 GiB | 8.6 | 8.8 | 9.3 | 10.1 | 11.8 | 15.2 |
| Q3_K_L | 7.38 GiB | 8.4 | 8.6 | 9.1 | 9.9 | 11.6 | 15.0 |
| Q3_K_M | 6.84 GiB | 7.9 | 8.1 | 8.5 | 9.4 | 11.1 | 14.4 |
| IQ3_M | 6.44 GiB | 7.5 | 7.7 | 8.1 | 9.0 | 10.7 | 14.0 |
| Q3_K_S | 6.20 GiB | 7.3 | 7.5 | 7.9 | 8.7 | 10.4 | 13.8 |
| Q2_K_L | 6.08 GiB | 7.1 | 7.4 | 7.8 | 8.6 | 10.3 | 13.7 |
| IQ3_XS | 5.94 GiB | 7.0 | 7.2 | 7.6 | 8.5 | 10.2 | 13.5 |
| Q2_K | 5.37 GiB | 6.4 | 6.6 | 7.1 | 7.9 | 9.6 | 13.0 |
| IQ2_M | 4.99 GiB | 6.0 | 6.3 | 6.7 | 7.5 | 9.2 | 12.6 |
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 131,072-token window rather than the full context.