Can I run Phi-3-mini-4k-instruct on a Radeon RX 6500 XT?
Not at these settings. No indexed quantization of Phi-3-mini-4k-instruct fits Radeon RX 6500 XT at any context we compute, with q8_0 KV. The smallest shipped quantization is 3.91 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◐ |
|---|---|---|---|---|---|---|---|
| F32 | 52.01 GiB | 53.7 | 54.5 | 56.1 | 59.3 | 65.7 | 78.4 |
| Q8_0 | 13.82 GiB | 15.5 | 16.3 | 17.9 | 21.1 | 27.5 | 40.2 |
| Q6_K | 10.67 GiB | 12.4 | 13.2 | 14.8 | 17.9 | 24.3 | 37.1 |
| Q5_K_M | 9.20 GiB | 10.9 | 11.7 | 13.3 | 16.5 | 22.9 | 35.6 |
| Q5_K_S | 8.96 GiB | 10.7 | 11.5 | 13.1 | 16.2 | 22.6 | 35.4 |
| Q4_K_M | 7.83 GiB | 9.5 | 10.3 | 11.9 | 15.1 | 21.5 | 34.2 |
| Q4_K_S | 7.41 GiB | 9.1 | 9.9 | 11.5 | 14.7 | 21.1 | 33.8 |
| IQ4_XS | 6.99 GiB | 8.7 | 9.5 | 11.1 | 14.3 | 20.6 | 33.4 |
| Q3_K_L | 6.84 GiB | 8.5 | 9.3 | 10.9 | 14.1 | 20.5 | 33.2 |
| Q3_K_M | 6.29 GiB | 8.0 | 8.8 | 10.4 | 13.6 | 19.9 | 32.7 |
| IQ3_M | 5.87 GiB | 7.6 | 8.4 | 10.0 | 13.1 | 19.5 | 32.3 |
| Q3_K_S | 5.65 GiB | 7.4 | 8.1 | 9.7 | 12.9 | 19.3 | 32.1 |
| IQ3_XS | 5.38 GiB | 7.1 | 7.9 | 9.5 | 12.7 | 19.0 | 31.8 |
| IQ3_XXS | 5.05 GiB | 6.7 | 7.5 | 9.1 | 12.3 | 18.7 | 31.5 |
| Q2_K | 4.85 GiB | 6.5 | 7.3 | 8.9 | 12.1 | 18.5 | 31.3 |
| IQ2_M | 4.45 GiB | 6.2 | 7.0 | 8.5 | 11.7 | 18.1 | 30.9 |
| IQ2_S | 4.11 GiB | 5.8 | 6.6 | 8.2 | 11.4 | 17.8 | 30.5 |
| IQ2_XS | 3.91 GiB | 5.6 | 6.4 | 8.0 | 11.2 | 17.6 | 30.3 |
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.