Can I run Mistral-Small-4-119B-2603 on a Radeon RX 6500 XT?
Not at these settings. No indexed quantization of Mistral-Small-4-119B-2603 fits Radeon RX 6500 XT at any context we compute, with q4_0 KV. The smallest shipped quantization is 25.04 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 | 221.65 GiB | 222.6 | 222.6 | 222.7 | 222.8 | 223.0 | 223.4 |
| Q8_0 | 117.79 GiB | 118.7 | 118.8 | 118.8 | 118.9 | 119.1 | 119.5 |
| Q6_K_L | 95.99 GiB | 96.9 | 97.0 | 97.0 | 97.1 | 97.3 | 97.7 |
| UD-Q6_K | 92.60 GiB | 93.6 | 93.6 | 93.6 | 93.7 | 93.9 | 94.3 |
| Q6_K | 90.96 GiB | 91.9 | 91.9 | 92.0 | 92.1 | 92.3 | 92.7 |
| UD-Q5_K_M | 83.04 GiB | 84.0 | 84.0 | 84.1 | 84.2 | 84.4 | 84.8 |
| Q5_K_L | 79.51 GiB | 80.5 | 80.5 | 80.5 | 80.6 | 80.8 | 81.2 |
| Q5_K_M | 79.20 GiB | 80.2 | 80.2 | 80.2 | 80.3 | 80.5 | 80.9 |
| UD-Q5_K_S | 77.82 GiB | 78.8 | 78.8 | 78.8 | 78.9 | 79.1 | 79.5 |
| Q5_K_S | 76.57 GiB | 77.5 | 77.6 | 77.6 | 77.7 | 77.9 | 78.3 |
| UD-Q4_K_M | 68.70 GiB | 69.7 | 69.7 | 69.7 | 69.8 | 70.0 | 70.4 |
| Q4_K_L | 68.02 GiB | 69.0 | 69.0 | 69.1 | 69.1 | 69.3 | 69.7 |
| Q4_K_M | 67.65 GiB | 68.6 | 68.6 | 68.7 | 68.8 | 69.0 | 69.4 |
| Q4_K_S | 65.08 GiB | 66.0 | 66.1 | 66.1 | 66.2 | 66.4 | 66.8 |
| UD-Q4_K_S | 64.70 GiB | 65.7 | 65.7 | 65.7 | 65.8 | 66.0 | 66.4 |
| Q4_0 | 63.06 GiB | 64.0 | 64.0 | 64.1 | 64.2 | 64.4 | 64.8 |
| IQ4_NL | 62.83 GiB | 63.8 | 63.8 | 63.9 | 64.0 | 64.2 | 64.6 |
| IQ4_XS | 59.41 GiB | 60.4 | 60.4 | 60.4 | 60.5 | 60.7 | 61.1 |
| UD-IQ4_NL | 55.13 GiB | 56.1 | 56.1 | 56.2 | 56.3 | 56.5 | 56.9 |
| UD-IQ4_XS | 54.13 GiB | 55.1 | 55.1 | 55.2 | 55.3 | 55.5 | 55.9 |
| IQ3_M | 53.05 GiB | 54.0 | 54.0 | 54.1 | 54.2 | 54.4 | 54.8 |
| Q3_K_L | 52.99 GiB | 53.9 | 54.0 | 54.0 | 54.1 | 54.3 | 54.7 |
| Q3_K_M | 50.82 GiB | 51.8 | 51.8 | 51.8 | 51.9 | 52.1 | 52.5 |
| UD-Q3_K_M | 50.64 GiB | 51.6 | 51.6 | 51.7 | 51.8 | 52.0 | 52.4 |
| Q3_K_S | 48.42 GiB | 49.4 | 49.4 | 49.5 | 49.6 | 49.8 | 50.1 |
| IQ3_XS | 48.20 GiB | 49.2 | 49.2 | 49.2 | 49.3 | 49.5 | 49.9 |
| IQ3_XXS | 46.44 GiB | 47.4 | 47.4 | 47.5 | 47.6 | 47.8 | 48.2 |
| UD-Q3_K_S | 46.23 GiB | 47.2 | 47.2 | 47.3 | 47.4 | 47.6 | 48.0 |
| UD-IQ3_S | 41.36 GiB | 42.3 | 42.3 | 42.4 | 42.5 | 42.7 | 43.1 |
| UD-IQ3_XXS | 39.86 GiB | 40.8 | 40.8 | 40.9 | 41.0 | 41.2 | 41.6 |
| Q2_K_L | 39.67 GiB | 40.6 | 40.7 | 40.7 | 40.8 | 41.0 | 41.4 |
| Q2_K | 39.18 GiB | 40.1 | 40.2 | 40.2 | 40.3 | 40.5 | 40.9 |
| IQ2_M | 38.77 GiB | 39.7 | 39.8 | 39.8 | 39.9 | 40.1 | 40.5 |
| IQ2_S | 35.52 GiB | 36.5 | 36.5 | 36.6 | 36.6 | 36.8 | 37.2 |
| UD-IQ2_M | 34.99 GiB | 36.0 | 36.0 | 36.0 | 36.1 | 36.3 | 36.7 |
| IQ2_XS | 34.48 GiB | 35.4 | 35.5 | 35.5 | 35.6 | 35.8 | 36.2 |
| UD-IQ2_XXS | 32.53 GiB | 33.5 | 33.5 | 33.6 | 33.7 | 33.9 | 34.2 |
| IQ2_XXS | 31.35 GiB | 32.3 | 32.3 | 32.4 | 32.5 | 32.7 | 33.1 |
| UD-IQ1_M | 30.09 GiB | 31.0 | 31.1 | 31.1 | 31.2 | 31.4 | 31.8 |
| IQ1_M | 27.51 GiB | 28.5 | 28.5 | 28.5 | 28.6 | 28.8 | 29.2 |
| IQ1_S | 25.04 GiB | 26.0 | 26.0 | 26.1 | 26.2 | 26.4 | 26.8 |
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, this model uses latent attention and allocates no V cache at all, so any formula reading num_key_value_heads overstates its cache by more than an order of magnitude.