Can I run GLM-5.1 on a Radeon RX 6500 XT?
Not at these settings. No indexed quantization of GLM-5.1 fits Radeon RX 6500 XT at any context we compute, with q4_0 KV. The smallest shipped quantization is 147.30 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 | 607.66 GiB | 608.7 | 608.8 | 609.0 | 609.4 | 610.2 | 611.7 |
| UD-Q6_K | 578.66 GiB | 579.7 | 579.8 | 580.0 | 580.4 | 581.2 | 582.7 |
| UD-Q5_K_M | 520.11 GiB | 521.2 | 521.2 | 521.4 | 521.8 | 522.6 | 524.1 |
| Q5_K_M | 501.62 GiB | 502.7 | 502.8 | 503.0 | 503.3 | 504.1 | 505.7 |
| UD-Q5_K_S | 489.83 GiB | 490.9 | 491.0 | 491.2 | 491.5 | 492.3 | 493.9 |
| Q5_K_S | 484.36 GiB | 485.4 | 485.5 | 485.7 | 486.1 | 486.9 | 488.4 |
| Q4_1 | 440.26 GiB | 441.3 | 441.4 | 441.6 | 442.0 | 442.7 | 444.3 |
| UD-Q4_K_M | 432.60 GiB | 433.6 | 433.7 | 433.9 | 434.3 | 435.1 | 436.6 |
| Q4_K_L | 428.77 GiB | 429.8 | 429.9 | 430.1 | 430.5 | 431.3 | 432.8 |
| Q4_K_M | 428.11 GiB | 429.2 | 429.2 | 429.4 | 429.8 | 430.6 | 432.1 |
| Q4_K_S | 411.10 GiB | 412.1 | 412.2 | 412.4 | 412.8 | 413.6 | 415.1 |
| UD-Q4_K_S | 404.10 GiB | 405.1 | 405.2 | 405.4 | 405.8 | 406.6 | 408.1 |
| Q4_0 | 398.24 GiB | 399.3 | 399.4 | 399.6 | 400.0 | 400.7 | 402.3 |
| IQ4_NL | 397.30 GiB | 398.3 | 398.4 | 398.6 | 399.0 | 399.8 | 401.3 |
| IQ4_XS | 375.69 GiB | 376.7 | 376.8 | 377.0 | 377.4 | 378.2 | 379.7 |
| UD-IQ4_NL | 343.54 GiB | 344.6 | 344.7 | 344.9 | 345.3 | 346.0 | 347.6 |
| UD-IQ4_XS | 336.51 GiB | 337.6 | 337.7 | 337.8 | 338.2 | 339.0 | 340.5 |
| IQ3_M | 335.52 GiB | 336.6 | 336.7 | 336.9 | 337.2 | 338.0 | 339.6 |
| Q3_K_L | 334.23 GiB | 335.3 | 335.4 | 335.6 | 335.9 | 336.7 | 338.3 |
| Q3_K_M | 320.96 GiB | 322.0 | 322.1 | 322.3 | 322.7 | 323.5 | 325.0 |
| IQ3_XS | 320.32 GiB | 321.4 | 321.5 | 321.7 | 322.0 | 322.8 | 324.3 |
| UD-Q3_K_M | 315.19 GiB | 316.2 | 316.3 | 316.5 | 316.9 | 317.7 | 319.2 |
| Q3_K_S | 305.77 GiB | 306.8 | 306.9 | 307.1 | 307.5 | 308.3 | 309.8 |
| IQ3_XXS | 293.34 GiB | 294.4 | 294.5 | 294.7 | 295.1 | 295.8 | 297.4 |
| UD-Q3_K_S | 291.94 GiB | 293.0 | 293.1 | 293.3 | 293.7 | 294.4 | 296.0 |
| UD-IQ3_S | 260.39 GiB | 261.4 | 261.5 | 261.7 | 262.1 | 262.9 | 264.4 |
| UD-IQ3_XXS | 249.84 GiB | 250.9 | 251.0 | 251.2 | 251.6 | 252.3 | 253.9 |
| Q2_K_L | 248.43 GiB | 249.5 | 249.6 | 249.8 | 250.2 | 250.9 | 252.5 |
| Q2_K | 247.56 GiB | 248.6 | 248.7 | 248.9 | 249.3 | 250.1 | 251.6 |
| IQ2_M | 237.40 GiB | 238.4 | 238.5 | 238.7 | 239.1 | 239.9 | 241.4 |
| UD-IQ2_M | 219.96 GiB | 221.0 | 221.1 | 221.3 | 221.7 | 222.5 | 224.0 |
| IQ2_S | 215.49 GiB | 216.5 | 216.6 | 216.8 | 217.2 | 218.0 | 219.5 |
| IQ2_XS | 211.04 GiB | 212.1 | 212.2 | 212.4 | 212.8 | 213.5 | 215.1 |
| UD-IQ2_XXS | 205.49 GiB | 206.5 | 206.6 | 206.8 | 207.2 | 208.0 | 209.5 |
| UD-IQ1_M | 191.42 GiB | 192.5 | 192.6 | 192.8 | 193.1 | 193.9 | 195.5 |
| IQ2_XXS | 189.92 GiB | 191.0 | 191.1 | 191.3 | 191.6 | 192.4 | 194.0 |
| IQ1_M | 163.93 GiB | 165.0 | 165.1 | 165.3 | 165.7 | 166.4 | 168.0 |
| IQ1_S | 147.30 GiB | 148.3 | 148.4 | 148.6 | 149.0 | 149.8 | 151.3 |
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, 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.