Can I run DeepSeek-R1 on a GeForce RTX 3080 Ti?
Not at these settings. No indexed quantization of DeepSeek-R1 fits GeForce RTX 3080 Ti at any context we compute, with q4_0 KV. The smallest shipped quantization is 124.38 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◐ |
|---|---|---|---|---|---|---|---|
| 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 | 512.97 GiB | 513.9 | 514.0 | 514.2 | 514.5 | 515.1 | 516.3 |
| Q5_K_M | 442.75 GiB | 443.7 | 443.8 | 443.9 | 444.2 | 444.8 | 446.0 |
| Q5_K_S | 430.10 GiB | 431.1 | 431.1 | 431.3 | 431.6 | 432.2 | 433.4 |
| Q4_1 | 391.10 GiB | 392.1 | 392.1 | 392.3 | 392.6 | 393.2 | 394.4 |
| Q4_K_M | 376.65 GiB | 377.6 | 377.7 | 377.8 | 378.1 | 378.7 | 379.9 |
| Q4_K_S | 353.90 GiB | 354.9 | 354.9 | 355.1 | 355.4 | 356.0 | 357.2 |
| Q4_0 | 353.00 GiB | 354.0 | 354.0 | 354.2 | 354.5 | 355.1 | 356.3 |
| IQ4_NL | 352.10 GiB | 353.1 | 353.1 | 353.3 | 353.6 | 354.2 | 355.4 |
| IQ4_XS | 332.60 GiB | 333.6 | 333.6 | 333.8 | 334.1 | 334.7 | 335.9 |
| Q3_K_L | 323.58 GiB | 324.5 | 324.6 | 324.8 | 325.1 | 325.7 | 326.9 |
| Q3_K_M | 297.28 GiB | 298.2 | 298.3 | 298.5 | 298.8 | 299.4 | 300.6 |
| IQ3_M | 272.03 GiB | 273.0 | 273.1 | 273.2 | 273.5 | 274.1 | 275.3 |
| Q3_K_S | 269.23 GiB | 270.2 | 270.3 | 270.4 | 270.7 | 271.3 | 272.5 |
| IQ3_XXS | 240.22 GiB | 241.2 | 241.2 | 241.4 | 241.7 | 242.3 | 243.5 |
| Q2_K_L | 228.11 GiB | 229.1 | 229.1 | 229.3 | 229.6 | 230.2 | 231.4 |
| Q2_K | 227.27 GiB | 228.2 | 228.3 | 228.4 | 228.8 | 229.4 | 230.6 |
| IQ2_M | 202.50 GiB | 203.5 | 203.5 | 203.7 | 204.0 | 204.6 | 205.8 |
| IQ2_S | 183.47 GiB | 184.4 | 184.5 | 184.6 | 184.9 | 185.6 | 186.8 |
| UD-IQ2_XXS | 182.69 GiB | 183.6 | 183.7 | 183.9 | 184.2 | 184.8 | 186.0 |
| IQ2_XS | 181.69 GiB | 182.6 | 182.7 | 182.9 | 183.2 | 183.8 | 185.0 |
| IQ2_XXS | 162.45 GiB | 163.4 | 163.5 | 163.6 | 163.9 | 164.5 | 165.7 |
| UD-IQ1_M | 157.32 GiB | 158.3 | 158.3 | 158.5 | 158.8 | 159.4 | 160.6 |
| IQ1_M | 138.66 GiB | 139.6 | 139.7 | 139.8 | 140.1 | 140.7 | 141.9 |
| UD-IQ1_S | 130.60 GiB | 131.6 | 131.6 | 131.8 | 132.1 | 132.7 | 133.9 |
| IQ1_S | 124.38 GiB | 125.3 | 125.4 | 125.6 | 125.9 | 126.5 | 127.7 |
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.