Can I run DeepSeek-R1-Distill-Qwen-14B-abliterated-v2 on a Radeon RX 6500 XT?
Not at these settings. No indexed quantization of DeepSeek-R1-Distill-Qwen-14B-abliterated-v2 fits Radeon RX 6500 XT at any context we compute, with q4_0 KV. The smallest shipped quantization is 3.36 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◐ |
|---|---|---|---|---|---|---|---|
| F16 | 27.52 GiB | 28.7 | 28.9 | 29.3 | 30.2 | 31.8 | 35.2 |
| Q8_0 | 14.62 GiB | 15.8 | 16.0 | 16.4 | 17.3 | 18.9 | 22.3 |
| I1-Q6_K | 11.29 GiB | 12.4 | 12.7 | 13.1 | 13.9 | 15.6 | 19.0 |
| Q6_K | 11.29 GiB | 12.4 | 12.7 | 13.1 | 13.9 | 15.6 | 19.0 |
| I1-Q5_K_M | 9.79 GiB | 10.9 | 11.2 | 11.6 | 12.4 | 14.1 | 17.5 |
| Q5_K_M | 9.79 GiB | 10.9 | 11.2 | 11.6 | 12.4 | 14.1 | 17.5 |
| I1-Q5_K_S | 9.56 GiB | 10.7 | 10.9 | 11.4 | 12.2 | 13.9 | 17.3 |
| Q5_K_S | 9.56 GiB | 10.7 | 10.9 | 11.4 | 12.2 | 13.9 | 17.3 |
| I1-Q4_1 | 8.75 GiB | 9.9 | 10.1 | 10.5 | 11.4 | 13.1 | 16.4 |
| I1-Q4_K_M | 8.37 GiB | 9.5 | 9.7 | 10.2 | 11.0 | 12.7 | 16.1 |
| Q4_K_M | 8.37 GiB | 9.5 | 9.7 | 10.2 | 11.0 | 12.7 | 16.1 |
| I1-Q4_K_S | 7.98 GiB | 9.1 | 9.4 | 9.8 | 10.6 | 12.3 | 15.7 |
| Q4_K_S | 7.98 GiB | 9.1 | 9.4 | 9.8 | 10.6 | 12.3 | 15.7 |
| I1-IQ4_NL | 7.96 GiB | 9.1 | 9.3 | 9.8 | 10.6 | 12.3 | 15.7 |
| I1-Q4_0 | 7.96 GiB | 9.1 | 9.3 | 9.7 | 10.6 | 12.3 | 15.7 |
| IQ4_XS | 7.62 GiB | 8.8 | 9.0 | 9.4 | 10.3 | 11.9 | 15.3 |
| I1-IQ4_XS | 7.56 GiB | 8.7 | 8.9 | 9.4 | 10.2 | 11.9 | 15.3 |
| I1-Q3_K_L | 7.38 GiB | 8.5 | 8.7 | 9.2 | 10.0 | 11.7 | 15.1 |
| Q3_K_L | 7.38 GiB | 8.5 | 8.7 | 9.2 | 10.0 | 11.7 | 15.1 |
| I1-Q3_K_M | 6.84 GiB | 8.0 | 8.2 | 8.6 | 9.5 | 11.2 | 14.5 |
| Q3_K_M | 6.84 GiB | 8.0 | 8.2 | 8.6 | 9.5 | 11.2 | 14.5 |
| I1-IQ3_M | 6.44 GiB | 7.6 | 7.8 | 8.2 | 9.1 | 10.8 | 14.1 |
| I1-IQ3_S | 6.23 GiB | 7.4 | 7.6 | 8.0 | 8.9 | 10.6 | 13.9 |
| I1-Q3_K_S | 6.20 GiB | 7.4 | 7.6 | 8.0 | 8.8 | 10.5 | 13.9 |
| Q3_K_S | 6.20 GiB | 7.4 | 7.6 | 8.0 | 8.8 | 10.5 | 13.9 |
| I1-IQ3_XS | 5.94 GiB | 7.1 | 7.3 | 7.7 | 8.6 | 10.3 | 13.6 |
| I1-IQ3_XXS | 5.54 GiB | 6.7 | 6.9 | 7.3 | 8.2 | 9.9 | 13.2 |
| I1-Q2_K | 5.37 GiB | 6.5 | 6.7 | 7.2 | 8.0 | 9.7 | 13.1 |
| Q2_K | 5.37 GiB | 6.5 | 6.7 | 7.2 | 8.0 | 9.7 | 13.1 |
| I1-Q2_K_S | 5.03 GiB | 6.2 | 6.4 | 6.8 | 7.7 | 9.3 | 12.7 |
| I1-IQ2_M | 4.99 GiB | 6.1 | 6.4 | 6.8 | 7.6 | 9.3 | 12.7 |
| I1-IQ2_S | 4.66 GiB | 5.8 | 6.0 | 6.5 | 7.3 | 9.0 | 12.4 |
| I1-IQ2_XS | 4.38 GiB | 5.5 | 5.7 | 6.2 | 7.0 | 8.7 | 12.1 |
| I1-IQ2_XXS | 4.02 GiB | 5.2 | 5.4 | 5.8 | 6.7 | 8.3 | 11.7 |
| I1-IQ1_M | 3.61 GiB | 4.8 | 5.0 | 5.4 | 6.2 | 7.9 | 11.3 |
| I1-IQ1_S | 3.36 GiB | 4.5 | 4.7 | 5.2 | 6.0 | 7.7 | 11.1 |
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.