Can I run GLM-5.2 on a GeForce RTX 5060 Ti?
Not at these settings. No indexed quantization of GLM-5.2 fits GeForce RTX 5060 Ti at any context we compute, with q8_0 KV. The smallest shipped quantization is 169.33 GiB in weights alone, against 14.88 GiB usable. CPU offload can still run it, slowly.
Every quantization at every context
| Quant | Weights● | 4K◐ | 8K◐ | 16K◐ | 32K◐ | 64K◐ | 128K◐ |
|---|---|---|---|---|---|---|---|
| BF16 | 1404.42 GiB | 1405.5 | 1405.6 | 1406.0 | 1406.7 | 1408.2 | 1411.1 |
| Q8_0 | 746.32 GiB | 747.4 | 747.5 | 747.9 | 748.6 | 750.1 | 753.0 |
| UD-Q6_K | 582.88 GiB | 583.9 | 584.1 | 584.5 | 585.2 | 586.6 | 589.6 |
| UD-Q5_K_M | 522.31 GiB | 523.3 | 523.5 | 523.9 | 524.6 | 526.1 | 529.0 |
| UD-Q5_K_S | 491.05 GiB | 492.1 | 492.3 | 492.6 | 493.4 | 494.8 | 497.7 |
| UD-Q4_K_M | 433.83 GiB | 434.9 | 435.0 | 435.4 | 436.1 | 437.6 | 440.5 |
| UD-Q4_K_S | 406.46 GiB | 407.5 | 407.7 | 408.0 | 408.8 | 410.2 | 413.1 |
| UD-IQ4_NL | 347.07 GiB | 348.1 | 348.3 | 348.6 | 349.4 | 350.8 | 353.7 |
| UD-IQ4_XS | 340.22 GiB | 341.3 | 341.4 | 341.8 | 342.5 | 344.0 | 346.9 |
| UD-Q3_K_M | 319.20 GiB | 320.2 | 320.4 | 320.8 | 321.5 | 323.0 | 325.9 |
| UD-IQ3_S | 287.44 GiB | 288.5 | 288.7 | 289.0 | 289.7 | 291.2 | 294.1 |
| UD-IQ3_XXS | 262.34 GiB | 263.4 | 263.6 | 263.9 | 264.6 | 266.1 | 269.0 |
| UD-IQ2_M | 222.19 GiB | 223.2 | 223.4 | 223.8 | 224.5 | 226.0 | 228.9 |
| UD-IQ2_XXS | 222.08 GiB | 223.1 | 223.3 | 223.7 | 224.4 | 225.8 | 228.8 |
| UD-IQ1_M | 215.35 GiB | 216.4 | 216.6 | 216.9 | 217.7 | 219.1 | 222.0 |
| UD-IQ1_S | 201.83 GiB | 202.9 | 203.0 | 203.4 | 204.1 | 205.6 | 208.5 |
| Q3_K_M | 169.33 GiB | 170.4 | 170.5 | 170.9 | 171.6 | 173.1 | 176.0 |
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.