Can I run DeepSeek-Coder-V2-Instruct on a GeForce RTX 2080 Ti?
Not at these settings. No indexed quantization of DeepSeek-Coder-V2-Instruct fits GeForce RTX 2080 Ti at any context we compute, with q4_0 KV. The smallest shipped quantization is 44.14 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◐ |
|---|---|---|---|---|---|---|---|
| F16 | 439.20 GiB | 440.1 | 440.2 | 440.3 | 440.6 | 441.2 | 442.4 |
| Q8_0 | 233.41 GiB | 234.3 | 234.4 | 234.5 | 234.8 | 235.4 | 236.6 |
| Q6_K | 180.25 GiB | 181.2 | 181.2 | 181.4 | 181.7 | 182.3 | 183.5 |
| Q5_K_M | 155.74 GiB | 156.7 | 156.7 | 156.9 | 157.2 | 157.8 | 159.0 |
| Q5_K | 155.74 GiB | 156.7 | 156.7 | 156.9 | 157.2 | 157.8 | 159.0 |
| Q5_0 | 151.16 GiB | 152.1 | 152.1 | 152.3 | 152.6 | 153.2 | 154.4 |
| Q5_K_S | 151.16 GiB | 152.1 | 152.1 | 152.3 | 152.6 | 153.2 | 154.4 |
| Q4_K | 132.67 GiB | 133.6 | 133.7 | 133.8 | 134.1 | 134.7 | 135.9 |
| Q4_K_M | 132.67 GiB | 133.6 | 133.7 | 133.8 | 134.1 | 134.7 | 135.9 |
| Q4_K_S | 124.68 GiB | 125.6 | 125.7 | 125.8 | 126.1 | 126.7 | 127.9 |
| IQ4_NL | 123.78 GiB | 124.7 | 124.8 | 124.9 | 125.2 | 125.8 | 127.0 |
| Q4_0 | 123.78 GiB | 124.7 | 124.8 | 124.9 | 125.2 | 125.8 | 127.0 |
| IQ4_XS | 116.94 GiB | 117.9 | 117.9 | 118.1 | 118.4 | 119.0 | 120.2 |
| Q3_K_L | 113.97 GiB | 114.9 | 115.0 | 115.1 | 115.4 | 116.0 | 117.2 |
| Q3_K | 104.93 GiB | 105.8 | 105.9 | 106.1 | 106.4 | 107.0 | 108.1 |
| Q3_K_M | 104.93 GiB | 105.8 | 105.9 | 106.1 | 106.4 | 107.0 | 108.1 |
| IQ3_M | 96.27 GiB | 97.2 | 97.3 | 97.4 | 97.7 | 98.3 | 99.5 |
| IQ3_S | 94.70 GiB | 95.6 | 95.7 | 95.8 | 96.1 | 96.7 | 97.9 |
| Q3_K_S | 94.70 GiB | 95.6 | 95.7 | 95.8 | 96.1 | 96.7 | 97.9 |
| IQ3_XS | 89.69 GiB | 90.6 | 90.7 | 90.8 | 91.1 | 91.7 | 92.9 |
| IQ3_XXS | 84.61 GiB | 85.5 | 85.6 | 85.7 | 86.0 | 86.6 | 87.8 |
| Q2_K_L | 81.44 GiB | 82.3 | 82.4 | 82.6 | 82.9 | 83.5 | 84.6 |
| Q2_K | 80.04 GiB | 81.0 | 81.0 | 81.2 | 81.5 | 82.1 | 83.3 |
| Q2_K_S | 74.13 GiB | 75.0 | 75.1 | 75.3 | 75.6 | 76.2 | 77.3 |
| IQ2_M | 71.64 GiB | 72.6 | 72.6 | 72.8 | 73.1 | 73.7 | 74.9 |
| IQ2_S | 65.07 GiB | 66.0 | 66.1 | 66.2 | 66.5 | 67.1 | 68.3 |
| IQ2_XS | 63.99 GiB | 64.9 | 65.0 | 65.1 | 65.4 | 66.0 | 67.2 |
| IQ2_XXS | 57.28 GiB | 58.2 | 58.3 | 58.4 | 58.7 | 59.3 | 60.5 |
| IQ1_M | 49.06 GiB | 50.0 | 50.1 | 50.2 | 50.5 | 51.1 | 52.3 |
| IQ1_S | 44.14 GiB | 45.1 | 45.1 | 45.3 | 45.6 | 46.2 | 47.3 |
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.