Can I run GLM-5 on a Radeon RX 6500 XT?
Not at these settings. No indexed quantization of GLM-5 fits Radeon RX 6500 XT at any context we compute, with q4_0 KV. The smallest shipped quantization is 164.05 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 | 1404.42 GiB | 1405.5 | 1405.6 | 1405.7 | 1406.1 | 1406.9 | 1408.4 |
| Q8_0 | 746.31 GiB | 747.4 | 747.5 | 747.6 | 748.0 | 748.8 | 750.3 |
| Q6_K | 576.84 GiB | 577.9 | 578.0 | 578.2 | 578.6 | 579.3 | 580.9 |
| Q5_K_M | 498.42 GiB | 499.5 | 499.6 | 499.7 | 500.1 | 500.9 | 502.4 |
| Q5_K_S | 484.04 GiB | 485.1 | 485.2 | 485.4 | 485.8 | 486.5 | 488.1 |
| Q4_1 | 440.24 GiB | 441.3 | 441.4 | 441.6 | 442.0 | 442.7 | 444.3 |
| Q4_K_M | 424.57 GiB | 425.6 | 425.7 | 425.9 | 426.3 | 427.1 | 428.6 |
| Q4_K_S | 398.95 GiB | 400.0 | 400.1 | 400.3 | 400.7 | 401.4 | 403.0 |
| Q4_0 | 397.81 GiB | 398.9 | 399.0 | 399.1 | 399.5 | 400.3 | 401.8 |
| IQ4_NL | 396.67 GiB | 397.7 | 397.8 | 398.0 | 398.4 | 399.2 | 400.7 |
| IQ4_XS | 375.23 GiB | 376.3 | 376.4 | 376.6 | 376.9 | 377.7 | 379.3 |
| Q3_K_M | 335.56 GiB | 336.6 | 336.7 | 336.9 | 337.3 | 338.1 | 339.6 |
| Q3_K_S | 303.87 GiB | 304.9 | 305.0 | 305.2 | 305.6 | 306.4 | 307.9 |
| UD-IQ3_XXS | 283.62 GiB | 284.7 | 284.8 | 285.0 | 285.3 | 286.1 | 287.7 |
| Q2_K_L | 257.21 GiB | 258.3 | 258.4 | 258.5 | 258.9 | 259.7 | 261.2 |
| Q2_K | 257.00 GiB | 258.0 | 258.1 | 258.3 | 258.7 | 259.5 | 261.0 |
| UD-IQ2_M | 237.27 GiB | 238.3 | 238.4 | 238.6 | 239.0 | 239.8 | 241.3 |
| UD-IQ2_XXS | 224.52 GiB | 225.6 | 225.7 | 225.9 | 226.2 | 227.0 | 228.6 |
| UD-IQ1_M | 208.48 GiB | 209.5 | 209.6 | 209.8 | 210.2 | 211.0 | 212.5 |
| UD-IQ1_S | 189.71 GiB | 190.8 | 190.9 | 191.0 | 191.4 | 192.2 | 193.7 |
| UD-TQ1_0 | 164.05 GiB | 165.1 | 165.2 | 165.4 | 165.8 | 166.5 | 168.1 |
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, 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.