Can I run GLM-4.6V on a GeForce RTX 5090 D V2?
Not at these settings. No indexed quantization of GLM-4.6V fits GeForce RTX 5090 D V2 at any context we compute, with f16 KV. The smallest shipped quantization is 33.46 GiB in weights alone, against 22.32 GiB usable. CPU offload can still run it, slowly.
Every quantization at every context
| Quant | Weights● | 4K◐ | 8K◐ | 16K◐ | 32K◐ | 64K◐ | 128K◐ |
|---|---|---|---|---|---|---|---|
| BF16 | 199.08 GiB | 200.6 | 201.3 | 202.8 | 205.7 | 211.4 | 222.9 |
| Q8_0 | 105.81 GiB | 107.4 | 108.1 | 109.5 | 112.4 | 118.1 | 129.6 |
| Q6_K_L | 89.58 GiB | 91.1 | 91.8 | 93.3 | 96.2 | 101.9 | 113.4 |
| Q6_K | 89.30 GiB | 90.9 | 91.6 | 93.0 | 95.9 | 101.6 | 113.1 |
| Q5_K_L | 75.96 GiB | 77.5 | 78.2 | 79.7 | 82.5 | 88.3 | 99.8 |
| Q5_K_M | 75.60 GiB | 77.2 | 77.9 | 79.3 | 82.2 | 87.9 | 99.4 |
| Q5_K_S | 70.79 GiB | 72.3 | 73.1 | 74.5 | 77.4 | 83.1 | 94.6 |
| Q4_K_L | 66.89 GiB | 68.4 | 69.2 | 70.6 | 73.5 | 79.2 | 90.7 |
| Q4_K_M | 66.46 GiB | 68.0 | 68.7 | 70.2 | 73.0 | 78.8 | 90.3 |
| Q4_1 | 62.66 GiB | 64.2 | 64.9 | 66.4 | 69.2 | 75.0 | 86.5 |
| Q4_K_S | 61.63 GiB | 63.2 | 63.9 | 65.3 | 68.2 | 74.0 | 85.5 |
| Q4_0 | 57.48 GiB | 59.0 | 59.7 | 61.2 | 64.1 | 69.8 | 81.3 |
| IQ4_NL | 56.83 GiB | 58.4 | 59.1 | 60.5 | 63.4 | 69.2 | 80.7 |
| IQ4_XS | 54.80 GiB | 56.4 | 57.1 | 58.5 | 61.4 | 67.1 | 78.6 |
| Q3_K_M | 51.64 GiB | 53.2 | 53.9 | 55.3 | 58.2 | 64.0 | 75.5 |
| Q3_K_L | 50.52 GiB | 52.1 | 52.8 | 54.2 | 57.1 | 62.8 | 74.3 |
| IQ3_M | 50.11 GiB | 51.7 | 52.4 | 53.8 | 56.7 | 62.4 | 73.9 |
| Q3_K_S | 48.19 GiB | 49.7 | 50.5 | 51.9 | 54.8 | 60.5 | 72.0 |
| UD-IQ3_XXS | 46.26 GiB | 47.8 | 48.5 | 50.0 | 52.8 | 58.6 | 70.1 |
| IQ3_XS | 45.85 GiB | 47.4 | 48.1 | 49.6 | 52.4 | 58.2 | 69.7 |
| IQ3_XXS | 44.99 GiB | 46.5 | 47.3 | 48.7 | 51.6 | 57.3 | 68.8 |
| Q2_K_L | 42.21 GiB | 43.8 | 44.5 | 45.9 | 48.8 | 54.5 | 66.0 |
| Q2_K | 41.64 GiB | 43.2 | 43.9 | 45.3 | 48.2 | 54.0 | 65.5 |
| IQ2_M | 40.26 GiB | 41.8 | 42.5 | 44.0 | 46.8 | 52.6 | 64.1 |
| UD-IQ2_M | 39.96 GiB | 41.5 | 42.2 | 43.7 | 46.5 | 52.3 | 63.8 |
| UD-IQ2_XXS | 38.37 GiB | 39.9 | 40.6 | 42.1 | 44.9 | 50.7 | 62.2 |
| IQ2_S | 37.82 GiB | 39.4 | 40.1 | 41.5 | 44.4 | 50.1 | 61.6 |
| IQ2_XS | 37.49 GiB | 39.0 | 39.8 | 41.2 | 44.1 | 49.8 | 61.3 |
| UD-IQ1_M | 35.93 GiB | 37.5 | 38.2 | 39.6 | 42.5 | 48.3 | 59.8 |
| IQ2_XXS | 35.11 GiB | 36.7 | 37.4 | 38.8 | 41.7 | 47.4 | 58.9 |
| UD-IQ1_S | 34.49 GiB | 36.0 | 36.8 | 38.2 | 41.1 | 46.8 | 58.3 |
| UD-TQ1_0 | 34.27 GiB | 35.8 | 36.5 | 38.0 | 40.8 | 46.6 | 58.1 |
| IQ1_M | 33.46 GiB | 35.0 | 35.7 | 37.2 | 40.0 | 45.8 | 57.3 |
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, the KV cache grows linearly with context and, past about 32K, becomes larger than the weights for many models.