Can I run dolphin-2.9.2-Phi-3-Medium on a Radeon RX 6500 XT?
Not at these settings. No indexed quantization of dolphin-2.9.2-Phi-3-Medium fits Radeon RX 6500 XT at any context we compute, with q4_0 KV. The smallest shipped quantization is 2.84 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◐ |
|---|---|---|---|---|---|---|---|
| BF16 | 26.00 GiB | 27.2 | 27.4 | 27.8 | 28.7 | 30.5 | 34.0 |
| Q8_0 | 13.82 GiB | 15.0 | 15.2 | 15.7 | 16.5 | 18.3 | 21.8 |
| Q6_K | 10.67 GiB | 11.8 | 12.1 | 12.5 | 13.4 | 15.1 | 18.7 |
| Q5_K | 9.20 GiB | 10.4 | 10.6 | 11.0 | 11.9 | 13.7 | 17.2 |
| Q5_K_S | 8.96 GiB | 10.1 | 10.4 | 10.8 | 11.7 | 13.4 | 17.0 |
| Q4_K | 7.83 GiB | 9.0 | 9.2 | 9.7 | 10.5 | 12.3 | 15.8 |
| Q4_K_S | 7.41 GiB | 8.6 | 8.8 | 9.2 | 10.1 | 11.9 | 15.4 |
| IQ4_NL | 7.39 GiB | 8.6 | 8.8 | 9.2 | 10.1 | 11.9 | 15.4 |
| IQ4_XS | 6.99 GiB | 8.2 | 8.4 | 8.8 | 9.7 | 11.5 | 15.0 |
| Q3_K_L | 6.84 GiB | 8.0 | 8.2 | 8.7 | 9.6 | 11.3 | 14.8 |
| Q3_K | 6.29 GiB | 7.5 | 7.7 | 8.1 | 9.0 | 10.8 | 14.3 |
| IQ3_M | 5.87 GiB | 7.0 | 7.3 | 7.7 | 8.6 | 10.3 | 13.9 |
| IQ3_S | 5.68 GiB | 6.9 | 7.1 | 7.5 | 8.4 | 10.2 | 13.7 |
| Q3_K_S | 5.65 GiB | 6.8 | 7.0 | 7.5 | 8.4 | 10.1 | 13.6 |
| IQ3_XS | 5.38 GiB | 6.6 | 6.8 | 7.2 | 8.1 | 9.9 | 13.4 |
| IQ3_XXS | 5.05 GiB | 6.2 | 6.4 | 6.9 | 7.8 | 9.5 | 13.0 |
| Q2_K | 4.85 GiB | 6.0 | 6.2 | 6.7 | 7.6 | 9.3 | 12.8 |
| Q2_K_S | 4.50 GiB | 5.7 | 5.9 | 6.3 | 7.2 | 9.0 | 12.5 |
| IQ2_M | 4.45 GiB | 5.6 | 5.9 | 6.3 | 7.2 | 8.9 | 12.4 |
| IQ2_S | 4.11 GiB | 5.3 | 5.5 | 5.9 | 6.8 | 8.6 | 12.1 |
| IQ2_XS | 3.91 GiB | 5.1 | 5.3 | 5.7 | 6.6 | 8.4 | 11.9 |
| IQ2_XXS | 3.53 GiB | 4.7 | 4.9 | 5.4 | 6.3 | 8.0 | 11.5 |
| IQ1_M | 3.10 GiB | 4.3 | 4.5 | 4.9 | 5.8 | 7.6 | 11.1 |
| IQ1_S | 2.84 GiB | 4.0 | 4.2 | 4.7 | 5.6 | 7.3 | 10.8 |
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 2,048-token window rather than the full context.