Can I run DeepSeek-V3.1 on a GeForce RTX 3080 Ti?
Not at these settings. No indexed quantization of DeepSeek-V3.1 fits GeForce RTX 3080 Ti at any context we compute, with q4_0 KV. The smallest shipped quantization is 137.32 GiB in weights alone, against 18.60 GiB usable. CPU offload can still run it, slowly.
Every quantization at every context
| Quant | Weights● | 4K◐ | 8K◐ | 16K◐ | 32K◐ | 64K◐ | 128K◐ |
|---|---|---|---|---|---|---|---|
| F16 | 1250.09 GiB | 1251.0 | 1251.1 | 1251.3 | 1251.6 | 1252.2 | 1253.4 |
| BF16 | 1250.09 GiB | 1251.0 | 1251.1 | 1251.3 | 1251.6 | 1252.2 | 1253.4 |
| Q8_0 | 664.30 GiB | 665.3 | 665.3 | 665.5 | 665.8 | 666.4 | 667.6 |
| Q6_K | 514.51 GiB | 515.5 | 515.5 | 515.7 | 516.0 | 516.6 | 517.8 |
| Q5_K_M | 445.49 GiB | 446.4 | 446.5 | 446.7 | 447.0 | 447.6 | 448.8 |
| Q5_K_S | 430.86 GiB | 431.8 | 431.9 | 432.0 | 432.3 | 432.9 | 434.2 |
| Q4_1 | 392.12 GiB | 393.1 | 393.1 | 393.3 | 393.6 | 394.2 | 395.4 |
| Q4_K_M | 381.12 GiB | 382.1 | 382.2 | 382.3 | 382.6 | 383.2 | 384.4 |
| Q4_K_S | 367.08 GiB | 368.0 | 368.1 | 368.3 | 368.6 | 369.2 | 370.4 |
| Q4_0 | 359.88 GiB | 360.8 | 360.9 | 361.1 | 361.4 | 362.0 | 363.2 |
| IQ4_NL | 354.35 GiB | 355.3 | 355.4 | 355.5 | 355.8 | 356.4 | 357.6 |
| IQ4_XS | 335.26 GiB | 336.2 | 336.3 | 336.4 | 336.7 | 337.3 | 338.5 |
| Q3_K_M | 298.44 GiB | 299.4 | 299.5 | 299.6 | 299.9 | 300.5 | 301.7 |
| Q3_K_L | 297.75 GiB | 298.7 | 298.8 | 298.9 | 299.2 | 299.8 | 301.0 |
| IQ3_M | 286.74 GiB | 287.7 | 287.8 | 287.9 | 288.2 | 288.8 | 290.0 |
| Q3_K_S | 273.21 GiB | 274.2 | 274.2 | 274.4 | 274.7 | 275.3 | 276.5 |
| UD-IQ3_XXS | 260.49 GiB | 261.4 | 261.5 | 261.7 | 262.0 | 262.6 | 263.8 |
| IQ3_XS | 258.11 GiB | 259.1 | 259.1 | 259.3 | 259.6 | 260.2 | 261.4 |
| IQ3_XXS | 249.25 GiB | 250.2 | 250.3 | 250.4 | 250.7 | 251.3 | 252.5 |
| Q2_K_L | 229.03 GiB | 230.0 | 230.1 | 230.2 | 230.5 | 231.1 | 232.3 |
| Q2_K | 228.82 GiB | 229.8 | 229.9 | 230.0 | 230.3 | 230.9 | 232.1 |
| UD-IQ2_M | 219.34 GiB | 220.3 | 220.4 | 220.5 | 220.8 | 221.4 | 222.6 |
| UD-IQ2_XXS | 210.37 GiB | 211.3 | 211.4 | 211.5 | 211.8 | 212.5 | 213.7 |
| IQ2_M | 200.27 GiB | 201.2 | 201.3 | 201.4 | 201.8 | 202.4 | 203.6 |
| UD-IQ1_M | 192.62 GiB | 193.6 | 193.7 | 193.8 | 194.1 | 194.7 | 195.9 |
| UD-IQ1_S | 179.11 GiB | 180.1 | 180.1 | 180.3 | 180.6 | 181.2 | 182.4 |
| IQ2_S | 176.61 GiB | 177.6 | 177.6 | 177.8 | 178.1 | 178.7 | 179.9 |
| IQ2_XS | 175.47 GiB | 176.4 | 176.5 | 176.6 | 176.9 | 177.6 | 178.8 |
| UD-TQ1_0 | 158.79 GiB | 159.7 | 159.8 | 160.0 | 160.3 | 160.9 | 162.1 |
| IQ2_XXS | 152.79 GiB | 153.7 | 153.8 | 154.0 | 154.3 | 154.9 | 156.1 |
| IQ1_M | 137.32 GiB | 138.3 | 138.4 | 138.5 | 138.8 | 139.4 | 140.6 |
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.