Can I run cogito-v2-preview-deepseek-671B-MoE on a Apple M5 Max?
Not at these settings. No indexed quantization of cogito-v2-preview-deepseek-671B-MoE fits Apple M5 Max at any context we compute, with q8_0 KV. The smallest shipped quantization is 150.73 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 | 1250.08 GiB | 1250.8 | 1251.0 | 1251.3 | 1251.8 | 1253.0 | 1255.3 |
| Q8_0 | 664.29 GiB | 665.1 | 665.2 | 665.5 | 666.1 | 667.2 | 669.5 |
| Q6_K | 513.20 GiB | 514.0 | 514.1 | 514.4 | 515.0 | 516.1 | 518.4 |
| Q5_K_M | 443.10 GiB | 443.9 | 444.0 | 444.3 | 444.9 | 446.0 | 448.3 |
| Q5_K_S | 430.45 GiB | 431.2 | 431.4 | 431.6 | 432.2 | 433.4 | 435.6 |
| Q4_1 | 391.51 GiB | 392.3 | 392.4 | 392.7 | 393.3 | 394.4 | 396.7 |
| Q4_K_M | 377.13 GiB | 377.9 | 378.0 | 378.3 | 378.9 | 380.0 | 382.3 |
| Q4_K_S | 354.38 GiB | 355.1 | 355.3 | 355.6 | 356.1 | 357.3 | 359.6 |
| Q4_0 | 353.48 GiB | 354.2 | 354.4 | 354.7 | 355.2 | 356.4 | 358.7 |
| IQ4_NL | 352.57 GiB | 353.3 | 353.5 | 353.8 | 354.3 | 355.5 | 357.8 |
| IQ4_XS | 333.13 GiB | 333.9 | 334.0 | 334.3 | 334.9 | 336.0 | 338.3 |
| Q3_K_M | 297.88 GiB | 298.6 | 298.8 | 299.1 | 299.6 | 300.8 | 303.1 |
| Q3_K_S | 269.83 GiB | 270.6 | 270.7 | 271.0 | 271.6 | 272.7 | 275.0 |
| Q2_K_L | 228.17 GiB | 228.9 | 229.1 | 229.4 | 229.9 | 231.1 | 233.4 |
| Q2_K | 227.97 GiB | 228.7 | 228.9 | 229.2 | 229.7 | 230.9 | 233.2 |
| UD-IQ2_M | 212.82 GiB | 213.6 | 213.7 | 214.0 | 214.6 | 215.7 | 218.0 |
| UD-IQ2_XXS | 201.88 GiB | 202.7 | 202.8 | 203.1 | 203.7 | 204.8 | 207.1 |
| UD-IQ1_M | 186.85 GiB | 187.6 | 187.8 | 188.0 | 188.6 | 189.8 | 192.0 |
| UD-IQ1_S | 173.12 GiB | 173.9 | 174.0 | 174.3 | 174.9 | 176.0 | 178.3 |
| UD-TQ1_0 | 150.73 GiB | 151.5 | 151.6 | 151.9 | 152.5 | 153.6 | 155.9 |
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.