Can I run DeepSeek-Prover-V2-671B on a GeForce RTX 2080 Ti?
Not at these settings. No indexed quantization of DeepSeek-Prover-V2-671B fits GeForce RTX 2080 Ti at any context we compute, with q8_0 KV. The smallest shipped quantization is 172.10 GiB in weights alone, against 10.23 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 | 1251.1 | 1251.3 | 1251.5 | 1252.1 | 1253.2 | 1255.5 |
| Q8_0 | 664.30 GiB | 665.3 | 665.5 | 665.7 | 666.3 | 667.5 | 669.7 |
| Q6_K | 513.09 GiB | 514.1 | 514.3 | 514.5 | 515.1 | 516.2 | 518.5 |
| Q5_K_M | 442.78 GiB | 443.8 | 443.9 | 444.2 | 444.8 | 445.9 | 448.2 |
| Q5_K_S | 394.82 GiB | 395.8 | 396.0 | 396.3 | 396.8 | 398.0 | 400.3 |
| Q4_1 | 391.10 GiB | 392.1 | 392.3 | 392.5 | 393.1 | 394.3 | 396.5 |
| Q4_K_M | 376.71 GiB | 377.7 | 377.9 | 378.2 | 378.7 | 379.9 | 382.2 |
| Q4_K_S | 353.96 GiB | 355.0 | 355.1 | 355.4 | 356.0 | 357.1 | 359.4 |
| Q4_0 | 353.24 GiB | 354.3 | 354.4 | 354.7 | 355.3 | 356.4 | 358.7 |
| IQ4_NL | 352.16 GiB | 353.2 | 353.3 | 353.6 | 354.2 | 355.3 | 357.6 |
| IQ4_XS | 332.72 GiB | 333.7 | 333.9 | 334.2 | 334.7 | 335.9 | 338.2 |
| Q3_K_M | 297.40 GiB | 298.4 | 298.6 | 298.8 | 299.4 | 300.6 | 302.8 |
| Q3_K_S | 269.35 GiB | 270.4 | 270.5 | 270.8 | 271.4 | 272.5 | 274.8 |
| UD-IQ3_XXS | 253.67 GiB | 254.7 | 254.8 | 255.1 | 255.7 | 256.8 | 259.1 |
| Q2_K_L | 227.64 GiB | 228.7 | 228.8 | 229.1 | 229.7 | 230.8 | 233.1 |
| Q2_K | 227.44 GiB | 228.5 | 228.6 | 228.9 | 229.5 | 230.6 | 232.9 |
| UD-IQ2_M | 212.09 GiB | 213.1 | 213.3 | 213.5 | 214.1 | 215.2 | 217.5 |
| UD-IQ2_XXS | 201.13 GiB | 202.1 | 202.3 | 202.6 | 203.1 | 204.3 | 206.6 |
| UD-IQ1_M | 186.06 GiB | 187.1 | 187.2 | 187.5 | 188.1 | 189.2 | 191.5 |
| UD-IQ1_S | 172.10 GiB | 173.1 | 173.3 | 173.5 | 174.1 | 175.3 | 177.5 |
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.