Can I run dolphincoder-starcoder2-15b on a Radeon RX 6500 XT?
Not at these settings. No indexed quantization of dolphincoder-starcoder2-15b fits Radeon RX 6500 XT at any context we compute, with q8_0 KV. The smallest shipped quantization is 3.31 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◐ |
|---|---|---|---|---|---|---|---|
| Q8_0 | 15.80 GiB | 17.0 | 17.1 | 17.5 | 18.1 | 19.4 | 22.1 |
| I1-Q6_K | 12.20 GiB | 13.4 | 13.5 | 13.9 | 14.5 | 15.8 | 18.5 |
| Q6_K | 12.20 GiB | 13.4 | 13.5 | 13.9 | 14.5 | 15.8 | 18.5 |
| I1-Q5_K_M | 10.65 GiB | 11.8 | 12.0 | 12.3 | 13.0 | 14.3 | 17.0 |
| Q5_K_M | 10.65 GiB | 11.8 | 12.0 | 12.3 | 13.0 | 14.3 | 17.0 |
| I1-Q5_K_S | 10.27 GiB | 11.4 | 11.6 | 11.9 | 12.6 | 13.9 | 16.6 |
| Q5_K_S | 10.27 GiB | 11.4 | 11.6 | 11.9 | 12.6 | 13.9 | 16.6 |
| I1-Q4_K_M | 9.18 GiB | 10.3 | 10.5 | 10.8 | 11.5 | 12.8 | 15.5 |
| Q4_K_M | 9.18 GiB | 10.3 | 10.5 | 10.8 | 11.5 | 12.8 | 15.5 |
| I1-Q4_K_S | 8.53 GiB | 9.7 | 9.9 | 10.2 | 10.9 | 12.2 | 14.8 |
| Q4_K_S | 8.53 GiB | 9.7 | 9.9 | 10.2 | 10.9 | 12.2 | 14.8 |
| I1-Q4_0 | 8.49 GiB | 9.6 | 9.8 | 10.1 | 10.8 | 12.1 | 14.8 |
| I1-Q3_K_L | 8.35 GiB | 9.5 | 9.7 | 10.0 | 10.7 | 12.0 | 14.7 |
| Q3_K_L | 8.35 GiB | 9.5 | 9.7 | 10.0 | 10.7 | 12.0 | 14.7 |
| IQ4_XS | 8.12 GiB | 9.3 | 9.4 | 9.8 | 10.4 | 11.8 | 14.4 |
| I1-IQ4_XS | 8.01 GiB | 9.2 | 9.3 | 9.7 | 10.3 | 11.7 | 14.3 |
| I1-Q3_K_M | 7.49 GiB | 8.7 | 8.8 | 9.1 | 9.8 | 11.1 | 13.8 |
| Q3_K_M | 7.49 GiB | 8.7 | 8.8 | 9.1 | 9.8 | 11.1 | 13.8 |
| I1-IQ3_M | 6.80 GiB | 8.0 | 8.1 | 8.5 | 9.1 | 10.5 | 13.1 |
| IQ3_M | 6.80 GiB | 8.0 | 8.1 | 8.5 | 9.1 | 10.5 | 13.1 |
| I1-IQ3_S | 6.52 GiB | 7.7 | 7.8 | 8.2 | 8.8 | 10.2 | 12.8 |
| IQ3_S | 6.52 GiB | 7.7 | 7.8 | 8.2 | 8.8 | 10.2 | 12.8 |
| I1-Q3_K_S | 6.51 GiB | 7.7 | 7.8 | 8.2 | 8.8 | 10.2 | 12.8 |
| Q3_K_S | 6.51 GiB | 7.7 | 7.8 | 8.2 | 8.8 | 10.2 | 12.8 |
| I1-IQ3_XS | 6.25 GiB | 7.4 | 7.6 | 7.9 | 8.6 | 9.9 | 12.6 |
| IQ3_XS | 6.25 GiB | 7.4 | 7.6 | 7.9 | 8.6 | 9.9 | 12.6 |
| I1-IQ3_XXS | 5.79 GiB | 6.9 | 7.1 | 7.4 | 8.1 | 9.4 | 12.1 |
| I1-Q2_K | 5.77 GiB | 6.9 | 7.1 | 7.4 | 8.1 | 9.4 | 12.1 |
| Q2_K | 5.77 GiB | 6.9 | 7.1 | 7.4 | 8.1 | 9.4 | 12.1 |
| I1-IQ2_M | 5.16 GiB | 6.3 | 6.5 | 6.8 | 7.5 | 8.8 | 11.5 |
| I1-IQ2_S | 4.79 GiB | 5.9 | 6.1 | 6.4 | 7.1 | 8.4 | 11.1 |
| I1-IQ2_XS | 4.49 GiB | 5.6 | 5.8 | 6.1 | 6.8 | 8.1 | 10.8 |
| I1-IQ2_XXS | 4.07 GiB | 5.2 | 5.4 | 5.7 | 6.4 | 7.7 | 10.4 |
| I1-IQ1_M | 3.60 GiB | 4.8 | 4.9 | 5.3 | 5.9 | 7.2 | 9.9 |
| I1-IQ1_S | 3.31 GiB | 4.5 | 4.6 | 5.0 | 5.6 | 7.0 | 9.6 |
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 4,096-token window rather than the full context.