Can I run Olmo-3-7B-Instruct on a Radeon RX 6500 XT?
Not at these settings. No indexed quantization of Olmo-3-7B-Instruct fits Radeon RX 6500 XT at any context we compute, with q8_0 KV. The smallest shipped quantization is 1.81 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 | 13.60 GiB | 15.6 | 16.0 | 16.5 | 17.5 | 19.7 | 23.9 |
| Q8_0 | 7.23 GiB | 9.2 | 9.6 | 10.1 | 11.2 | 13.3 | 17.5 |
| Q6_K_L | 5.77 GiB | 7.8 | 8.1 | 8.7 | 9.7 | 11.8 | 16.1 |
| Q6_K | 5.58 GiB | 7.6 | 7.9 | 8.5 | 9.5 | 11.7 | 15.9 |
| Q5_K_L | 5.09 GiB | 7.1 | 7.4 | 8.0 | 9.0 | 11.2 | 15.4 |
| Q5_K_M | 4.85 GiB | 6.8 | 7.2 | 7.7 | 8.8 | 10.9 | 15.2 |
| Q5_K_S | 4.73 GiB | 6.7 | 7.1 | 7.6 | 8.7 | 10.8 | 15.1 |
| Q4_K_L | 4.45 GiB | 6.4 | 6.8 | 7.3 | 8.4 | 10.5 | 14.8 |
| Q4_1 | 4.33 GiB | 6.3 | 6.7 | 7.2 | 8.3 | 10.4 | 14.7 |
| Q4_K_M | 4.16 GiB | 6.2 | 6.5 | 7.0 | 8.1 | 10.2 | 14.5 |
| Q4_K_S | 3.96 GiB | 5.9 | 6.3 | 6.8 | 7.9 | 10.0 | 14.3 |
| Q4_0 | 3.94 GiB | 5.9 | 6.3 | 6.8 | 7.9 | 10.0 | 14.3 |
| IQ4_NL | 3.93 GiB | 5.9 | 6.3 | 6.8 | 7.9 | 10.0 | 14.2 |
| IQ4_XS | 3.74 GiB | 5.7 | 6.1 | 6.6 | 7.7 | 9.8 | 14.1 |
| Q3_K_L | 3.68 GiB | 5.7 | 6.0 | 6.6 | 7.6 | 9.8 | 14.0 |
| Q3_K_M | 3.40 GiB | 5.4 | 5.8 | 6.3 | 7.3 | 9.5 | 13.7 |
| IQ3_M | 3.23 GiB | 5.2 | 5.6 | 6.1 | 7.2 | 9.3 | 13.6 |
| Q3_K_S | 3.08 GiB | 5.1 | 5.4 | 6.0 | 7.0 | 9.1 | 13.4 |
| Q2_K_L | 3.04 GiB | 5.0 | 5.4 | 5.9 | 7.0 | 9.1 | 13.4 |
| IQ3_XS | 2.93 GiB | 4.9 | 5.3 | 5.8 | 6.9 | 9.0 | 13.3 |
| UD-IQ3_XXS | 2.74 GiB | 4.7 | 5.1 | 5.6 | 6.7 | 8.8 | 13.1 |
| IQ3_XXS | 2.70 GiB | 4.7 | 5.1 | 5.6 | 6.7 | 8.8 | 13.0 |
| Q2_K | 2.66 GiB | 4.7 | 5.0 | 5.5 | 6.6 | 8.7 | 13.0 |
| UD-IQ2_M | 2.53 GiB | 4.5 | 4.9 | 5.4 | 6.5 | 8.6 | 12.9 |
| IQ2_M | 2.49 GiB | 4.5 | 4.8 | 5.4 | 6.4 | 8.6 | 12.8 |
| UD-IQ2_XXS | 2.10 GiB | 4.1 | 4.5 | 5.0 | 6.0 | 8.2 | 12.4 |
| UD-IQ1_M | 1.90 GiB | 3.9 | 4.3 | 4.8 | 5.8 | 8.0 | 12.2 |
| UD-IQ1_S | 1.81 GiB | 3.8 | 4.2 | 4.7 | 5.8 | 7.9 | 12.1 |
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 4,096-token window rather than the full context.