Can I run DeepSeek-V3-0324 on a Apple M3 Ultra?
Not at these settings. No indexed quantization of DeepSeek-V3-0324 fits Apple M3 Ultra at any context we compute, with q8_0 KV. The smallest shipped quantization is 124.38 GiB in weights alone, against 66.96 GiB usable. CPU offload can still run it, slowly.
Every quantization at every context
| Quant | Weights● | 4K◐ | 8K◐ | 16K◐ | 32K◐ | 64K◐ | 128K◐ |
|---|---|---|---|---|---|---|---|
| BF16 | 1250.09 GiB | 1250.9 | 1251.0 | 1251.3 | 1251.9 | 1253.0 | 1255.3 |
| Q8_0 | 664.30 GiB | 665.1 | 665.2 | 665.5 | 666.1 | 667.2 | 669.5 |
| Q6_K | 513.09 GiB | 513.9 | 514.0 | 514.3 | 514.9 | 516.0 | 518.3 |
| Q5_K_M | 442.93 GiB | 443.7 | 443.8 | 444.1 | 444.7 | 445.8 | 448.1 |
| Q5_K_S | 430.10 GiB | 430.9 | 431.0 | 431.3 | 431.9 | 433.0 | 435.3 |
| Q4_1 | 391.10 GiB | 391.9 | 392.0 | 392.3 | 392.9 | 394.0 | 396.3 |
| Q4_K_M | 376.89 GiB | 377.7 | 377.8 | 378.1 | 378.7 | 379.8 | 382.1 |
| Q4_K_S | 353.90 GiB | 354.7 | 354.8 | 355.1 | 355.7 | 356.8 | 359.1 |
| Q4_0 | 353.00 GiB | 353.8 | 353.9 | 354.2 | 354.8 | 355.9 | 358.2 |
| IQ4_NL | 352.34 GiB | 353.1 | 353.3 | 353.5 | 354.1 | 355.2 | 357.5 |
| IQ4_XS | 332.90 GiB | 333.7 | 333.8 | 334.1 | 334.7 | 335.8 | 338.1 |
| Q3_K_M | 297.58 GiB | 298.3 | 298.5 | 298.8 | 299.3 | 300.5 | 302.8 |
| UD-IQ3_XXS | 254.55 GiB | 255.3 | 255.5 | 255.7 | 256.3 | 257.5 | 259.7 |
| IQ3_XXS | 240.22 GiB | 241.0 | 241.1 | 241.4 | 242.0 | 243.1 | 245.4 |
| Q2_K_L | 228.11 GiB | 228.9 | 229.0 | 229.3 | 229.9 | 231.0 | 233.3 |
| Q2_K | 227.62 GiB | 228.4 | 228.5 | 228.8 | 229.4 | 230.5 | 232.8 |
| UD-IQ2_M | 212.09 GiB | 212.9 | 213.0 | 213.3 | 213.9 | 215.0 | 217.3 |
| UD-IQ2_XXS | 203.64 GiB | 204.4 | 204.5 | 204.8 | 205.4 | 206.5 | 208.8 |
| IQ2_M | 202.50 GiB | 203.3 | 203.4 | 203.7 | 204.3 | 205.4 | 207.7 |
| UD-IQ1_M | 186.94 GiB | 187.7 | 187.9 | 188.1 | 188.7 | 189.8 | 192.1 |
| IQ2_S | 183.47 GiB | 184.2 | 184.4 | 184.7 | 185.2 | 186.4 | 188.7 |
| UD-IQ1_S | 173.45 GiB | 174.2 | 174.4 | 174.6 | 175.2 | 176.4 | 178.6 |
| IQ2_XXS | 162.45 GiB | 163.2 | 163.4 | 163.6 | 164.2 | 165.4 | 167.6 |
| IQ1_M | 138.66 GiB | 139.4 | 139.6 | 139.9 | 140.4 | 141.6 | 143.8 |
| IQ1_S | 124.38 GiB | 125.2 | 125.3 | 125.6 | 126.2 | 127.3 | 129.6 |
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.