Can I run cogito-v2-preview-deepseek-671B-MoE on a Apple M5 Pro?
Not at these settings. No indexed quantization of cogito-v2-preview-deepseek-671B-MoE fits Apple M5 Pro at any context we compute, with q4_0 KV. The smallest shipped quantization is 150.73 GiB in weights alone, against 33.48 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 | 1250.9 | 1251.0 | 1251.3 | 1251.9 | 1253.1 |
| Q8_0 | 664.29 GiB | 665.0 | 665.1 | 665.2 | 665.5 | 666.1 | 667.3 |
| Q6_K | 513.20 GiB | 513.9 | 514.0 | 514.1 | 514.4 | 515.0 | 516.2 |
| Q5_K_M | 443.10 GiB | 443.8 | 443.9 | 444.0 | 444.3 | 444.9 | 446.1 |
| Q5_K_S | 430.45 GiB | 431.2 | 431.2 | 431.4 | 431.7 | 432.3 | 433.5 |
| Q4_1 | 391.51 GiB | 392.2 | 392.3 | 392.4 | 392.7 | 393.3 | 394.6 |
| Q4_K_M | 377.13 GiB | 377.8 | 377.9 | 378.1 | 378.4 | 379.0 | 380.2 |
| Q4_K_S | 354.38 GiB | 355.1 | 355.2 | 355.3 | 355.6 | 356.2 | 357.4 |
| Q4_0 | 353.48 GiB | 354.2 | 354.3 | 354.4 | 354.7 | 355.3 | 356.5 |
| IQ4_NL | 352.57 GiB | 353.3 | 353.4 | 353.5 | 353.8 | 354.4 | 355.6 |
| IQ4_XS | 333.13 GiB | 333.8 | 333.9 | 334.1 | 334.4 | 335.0 | 336.2 |
| Q3_K_M | 297.88 GiB | 298.6 | 298.7 | 298.8 | 299.1 | 299.7 | 300.9 |
| Q3_K_S | 269.83 GiB | 270.5 | 270.6 | 270.8 | 271.1 | 271.7 | 272.9 |
| Q2_K_L | 228.17 GiB | 228.9 | 228.9 | 229.1 | 229.4 | 230.0 | 231.2 |
| Q2_K | 227.97 GiB | 228.7 | 228.7 | 228.9 | 229.2 | 229.8 | 231.0 |
| UD-IQ2_M | 212.82 GiB | 213.5 | 213.6 | 213.8 | 214.1 | 214.7 | 215.9 |
| UD-IQ2_XXS | 201.88 GiB | 202.6 | 202.7 | 202.8 | 203.1 | 203.7 | 204.9 |
| UD-IQ1_M | 186.85 GiB | 187.6 | 187.6 | 187.8 | 188.1 | 188.7 | 189.9 |
| UD-IQ1_S | 173.12 GiB | 173.8 | 173.9 | 174.1 | 174.4 | 175.0 | 176.2 |
| UD-TQ1_0 | 150.73 GiB | 151.4 | 151.5 | 151.7 | 152.0 | 152.6 | 153.8 |
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.