Can I run GLM-5.1 on a Apple M5 Max?
Not at these settings. No indexed quantization of GLM-5.1 fits Apple M5 Max at any context we compute, with q8_0 KV. The smallest shipped quantization is 147.30 GiB in weights alone, against 25.11 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 | 1405.2 | 1405.4 | 1405.7 | 1406.5 | 1407.9 | 1410.8 |
| Q8_0 | 746.31 GiB | 747.1 | 747.3 | 747.6 | 748.4 | 749.8 | 752.7 |
| Q6_K | 607.66 GiB | 608.4 | 608.6 | 609.0 | 609.7 | 611.2 | 614.1 |
| UD-Q6_K | 578.66 GiB | 579.4 | 579.6 | 580.0 | 580.7 | 582.2 | 585.1 |
| UD-Q5_K_M | 520.11 GiB | 520.9 | 521.1 | 521.4 | 522.2 | 523.6 | 526.5 |
| Q5_K_M | 501.62 GiB | 502.4 | 502.6 | 502.9 | 503.7 | 505.1 | 508.0 |
| UD-Q5_K_S | 489.83 GiB | 490.6 | 490.8 | 491.2 | 491.9 | 493.3 | 496.3 |
| Q5_K_S | 484.36 GiB | 485.1 | 485.3 | 485.7 | 486.4 | 487.9 | 490.8 |
| Q4_1 | 440.26 GiB | 441.0 | 441.2 | 441.6 | 442.3 | 443.8 | 446.7 |
| UD-Q4_K_M | 432.60 GiB | 433.4 | 433.6 | 433.9 | 434.7 | 436.1 | 439.0 |
| Q4_K_L | 428.77 GiB | 429.5 | 429.7 | 430.1 | 430.8 | 432.3 | 435.2 |
| Q4_K_M | 428.11 GiB | 428.9 | 429.1 | 429.4 | 430.2 | 431.6 | 434.5 |
| Q4_K_S | 411.10 GiB | 411.9 | 412.1 | 412.4 | 413.2 | 414.6 | 417.5 |
| UD-Q4_K_S | 404.10 GiB | 404.9 | 405.1 | 405.4 | 406.2 | 407.6 | 410.5 |
| Q4_0 | 398.24 GiB | 399.0 | 399.2 | 399.6 | 400.3 | 401.8 | 404.7 |
| IQ4_NL | 397.30 GiB | 398.1 | 398.3 | 398.6 | 399.4 | 400.8 | 403.7 |
| IQ4_XS | 375.69 GiB | 376.5 | 376.7 | 377.0 | 377.7 | 379.2 | 382.1 |
| UD-IQ4_NL | 343.54 GiB | 344.3 | 344.5 | 344.9 | 345.6 | 347.1 | 350.0 |
| UD-IQ4_XS | 336.51 GiB | 337.3 | 337.5 | 337.8 | 338.6 | 340.0 | 342.9 |
| IQ3_M | 335.52 GiB | 336.3 | 336.5 | 336.8 | 337.6 | 339.0 | 341.9 |
| Q3_K_L | 334.23 GiB | 335.0 | 335.2 | 335.6 | 336.3 | 337.7 | 340.7 |
| Q3_K_M | 320.96 GiB | 321.7 | 321.9 | 322.3 | 323.0 | 324.5 | 327.4 |
| IQ3_XS | 320.32 GiB | 321.1 | 321.3 | 321.6 | 322.4 | 323.8 | 326.7 |
| UD-Q3_K_M | 315.19 GiB | 316.0 | 316.2 | 316.5 | 317.2 | 318.7 | 321.6 |
| Q3_K_S | 305.77 GiB | 306.5 | 306.7 | 307.1 | 307.8 | 309.3 | 312.2 |
| IQ3_XXS | 293.34 GiB | 294.1 | 294.3 | 294.7 | 295.4 | 296.9 | 299.8 |
| UD-Q3_K_S | 291.94 GiB | 292.7 | 292.9 | 293.3 | 294.0 | 295.5 | 298.4 |
| UD-IQ3_S | 260.39 GiB | 261.2 | 261.4 | 261.7 | 262.4 | 263.9 | 266.8 |
| UD-IQ3_XXS | 249.84 GiB | 250.6 | 250.8 | 251.2 | 251.9 | 253.4 | 256.3 |
| Q2_K_L | 248.43 GiB | 249.2 | 249.4 | 249.8 | 250.5 | 251.9 | 254.9 |
| Q2_K | 247.56 GiB | 248.3 | 248.5 | 248.9 | 249.6 | 251.1 | 254.0 |
| IQ2_M | 237.40 GiB | 238.2 | 238.4 | 238.7 | 239.5 | 240.9 | 243.8 |
| UD-IQ2_M | 219.96 GiB | 220.7 | 220.9 | 221.3 | 222.0 | 223.5 | 226.4 |
| IQ2_S | 215.49 GiB | 216.3 | 216.5 | 216.8 | 217.5 | 219.0 | 221.9 |
| IQ2_XS | 211.04 GiB | 211.8 | 212.0 | 212.4 | 213.1 | 214.5 | 217.5 |
| UD-IQ2_XXS | 205.49 GiB | 206.3 | 206.4 | 206.8 | 207.5 | 209.0 | 211.9 |
| UD-IQ1_M | 191.42 GiB | 192.2 | 192.4 | 192.8 | 193.5 | 194.9 | 197.8 |
| IQ2_XXS | 189.92 GiB | 190.7 | 190.9 | 191.3 | 192.0 | 193.4 | 196.3 |
| IQ1_M | 163.93 GiB | 164.7 | 164.9 | 165.3 | 166.0 | 167.4 | 170.4 |
| IQ1_S | 147.30 GiB | 148.1 | 148.3 | 148.6 | 149.4 | 150.8 | 153.7 |
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.