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 q4_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 | 1405.1 | 1405.2 | 1405.4 | 1405.8 | 1406.6 | 1408.1 |
| Q8_0 | 746.31 GiB | 747.0 | 747.1 | 747.3 | 747.7 | 748.5 | 750.0 |
| Q6_K | 607.66 GiB | 608.4 | 608.5 | 608.6 | 609.0 | 609.8 | 611.3 |
| UD-Q6_K | 578.66 GiB | 579.4 | 579.5 | 579.6 | 580.0 | 580.8 | 582.3 |
| UD-Q5_K_M | 520.11 GiB | 520.8 | 520.9 | 521.1 | 521.5 | 522.2 | 523.8 |
| Q5_K_M | 501.62 GiB | 502.3 | 502.4 | 502.6 | 503.0 | 503.8 | 505.3 |
| UD-Q5_K_S | 489.83 GiB | 490.5 | 490.6 | 490.8 | 491.2 | 492.0 | 493.5 |
| Q5_K_S | 484.36 GiB | 485.1 | 485.2 | 485.3 | 485.7 | 486.5 | 488.0 |
| Q4_1 | 440.26 GiB | 441.0 | 441.1 | 441.2 | 441.6 | 442.4 | 443.9 |
| UD-Q4_K_M | 432.60 GiB | 433.3 | 433.4 | 433.6 | 434.0 | 434.7 | 436.3 |
| Q4_K_L | 428.77 GiB | 429.5 | 429.6 | 429.7 | 430.1 | 430.9 | 432.4 |
| Q4_K_M | 428.11 GiB | 428.8 | 428.9 | 429.1 | 429.5 | 430.2 | 431.8 |
| Q4_K_S | 411.10 GiB | 411.8 | 411.9 | 412.1 | 412.5 | 413.2 | 414.8 |
| UD-Q4_K_S | 404.10 GiB | 404.8 | 404.9 | 405.1 | 405.5 | 406.2 | 407.8 |
| Q4_0 | 398.24 GiB | 398.9 | 399.0 | 399.2 | 399.6 | 400.4 | 401.9 |
| IQ4_NL | 397.30 GiB | 398.0 | 398.1 | 398.3 | 398.7 | 399.4 | 401.0 |
| IQ4_XS | 375.69 GiB | 376.4 | 376.5 | 376.7 | 377.1 | 377.8 | 379.4 |
| UD-IQ4_NL | 343.54 GiB | 344.2 | 344.3 | 344.5 | 344.9 | 345.7 | 347.2 |
| UD-IQ4_XS | 336.51 GiB | 337.2 | 337.3 | 337.5 | 337.9 | 338.7 | 340.2 |
| IQ3_M | 335.52 GiB | 336.2 | 336.3 | 336.5 | 336.9 | 337.7 | 339.2 |
| Q3_K_L | 334.23 GiB | 334.9 | 335.0 | 335.2 | 335.6 | 336.4 | 337.9 |
| Q3_K_M | 320.96 GiB | 321.7 | 321.8 | 321.9 | 322.3 | 323.1 | 324.6 |
| IQ3_XS | 320.32 GiB | 321.0 | 321.1 | 321.3 | 321.7 | 322.5 | 324.0 |
| UD-Q3_K_M | 315.19 GiB | 315.9 | 316.0 | 316.2 | 316.6 | 317.3 | 318.9 |
| Q3_K_S | 305.77 GiB | 306.5 | 306.6 | 306.8 | 307.1 | 307.9 | 309.5 |
| IQ3_XXS | 293.34 GiB | 294.0 | 294.1 | 294.3 | 294.7 | 295.5 | 297.0 |
| UD-Q3_K_S | 291.94 GiB | 292.6 | 292.7 | 292.9 | 293.3 | 294.1 | 295.6 |
| UD-IQ3_S | 260.39 GiB | 261.1 | 261.2 | 261.4 | 261.8 | 262.5 | 264.1 |
| UD-IQ3_XXS | 249.84 GiB | 250.5 | 250.6 | 250.8 | 251.2 | 252.0 | 253.5 |
| Q2_K_L | 248.43 GiB | 249.1 | 249.2 | 249.4 | 249.8 | 250.6 | 252.1 |
| Q2_K | 247.56 GiB | 248.3 | 248.4 | 248.5 | 248.9 | 249.7 | 251.2 |
| IQ2_M | 237.40 GiB | 238.1 | 238.2 | 238.4 | 238.8 | 239.5 | 241.1 |
| UD-IQ2_M | 219.96 GiB | 220.7 | 220.8 | 220.9 | 221.3 | 222.1 | 223.6 |
| IQ2_S | 215.49 GiB | 216.2 | 216.3 | 216.5 | 216.9 | 217.6 | 219.2 |
| IQ2_XS | 211.04 GiB | 211.7 | 211.8 | 212.0 | 212.4 | 213.2 | 214.7 |
| UD-IQ2_XXS | 205.49 GiB | 206.2 | 206.3 | 206.5 | 206.9 | 207.6 | 209.2 |
| UD-IQ1_M | 191.42 GiB | 192.1 | 192.2 | 192.4 | 192.8 | 193.6 | 195.1 |
| IQ2_XXS | 189.92 GiB | 190.6 | 190.7 | 190.9 | 191.3 | 192.1 | 193.6 |
| IQ1_M | 163.93 GiB | 164.6 | 164.7 | 164.9 | 165.3 | 166.1 | 167.6 |
| IQ1_S | 147.30 GiB | 148.0 | 148.1 | 148.3 | 148.7 | 149.4 | 151.0 |
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.