Can I run DeepSeek-V3.1 on a Radeon RX 6500 XT?
Not at these settings. No indexed quantization of DeepSeek-V3.1 fits Radeon RX 6500 XT at any context we compute, with q4_0 KV. The smallest shipped quantization is 137.32 GiB in weights alone, against 3.72 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 | 1251.1 | 1251.2 | 1251.4 | 1251.7 | 1252.3 | 1253.5 |
| BF16 | 1250.09 GiB | 1251.1 | 1251.2 | 1251.4 | 1251.7 | 1252.3 | 1253.5 |
| Q8_0 | 664.30 GiB | 665.4 | 665.4 | 665.6 | 665.9 | 666.5 | 667.7 |
| Q6_K | 514.51 GiB | 515.6 | 515.6 | 515.8 | 516.1 | 516.7 | 517.9 |
| Q5_K_M | 445.49 GiB | 446.5 | 446.6 | 446.8 | 447.1 | 447.7 | 448.9 |
| Q5_K_S | 430.86 GiB | 431.9 | 432.0 | 432.1 | 432.4 | 433.0 | 434.3 |
| Q4_1 | 392.12 GiB | 393.2 | 393.2 | 393.4 | 393.7 | 394.3 | 395.5 |
| Q4_K_M | 381.12 GiB | 382.2 | 382.3 | 382.4 | 382.7 | 383.3 | 384.5 |
| Q4_K_S | 367.08 GiB | 368.1 | 368.2 | 368.4 | 368.7 | 369.3 | 370.5 |
| Q4_0 | 359.88 GiB | 360.9 | 361.0 | 361.2 | 361.5 | 362.1 | 363.3 |
| IQ4_NL | 354.35 GiB | 355.4 | 355.5 | 355.6 | 355.9 | 356.5 | 357.7 |
| IQ4_XS | 335.26 GiB | 336.3 | 336.4 | 336.5 | 336.8 | 337.4 | 338.6 |
| Q3_K_M | 298.44 GiB | 299.5 | 299.6 | 299.7 | 300.0 | 300.6 | 301.8 |
| Q3_K_L | 297.75 GiB | 298.8 | 298.9 | 299.0 | 299.3 | 299.9 | 301.1 |
| IQ3_M | 286.74 GiB | 287.8 | 287.9 | 288.0 | 288.3 | 288.9 | 290.1 |
| Q3_K_S | 273.21 GiB | 274.3 | 274.3 | 274.5 | 274.8 | 275.4 | 276.6 |
| UD-IQ3_XXS | 260.49 GiB | 261.5 | 261.6 | 261.8 | 262.1 | 262.7 | 263.9 |
| IQ3_XS | 258.11 GiB | 259.2 | 259.2 | 259.4 | 259.7 | 260.3 | 261.5 |
| IQ3_XXS | 249.25 GiB | 250.3 | 250.4 | 250.5 | 250.8 | 251.4 | 252.6 |
| Q2_K_L | 229.03 GiB | 230.1 | 230.2 | 230.3 | 230.6 | 231.2 | 232.4 |
| Q2_K | 228.82 GiB | 229.9 | 230.0 | 230.1 | 230.4 | 231.0 | 232.2 |
| UD-IQ2_M | 219.34 GiB | 220.4 | 220.5 | 220.6 | 220.9 | 221.5 | 222.7 |
| UD-IQ2_XXS | 210.37 GiB | 211.4 | 211.5 | 211.6 | 211.9 | 212.6 | 213.8 |
| IQ2_M | 200.27 GiB | 201.3 | 201.4 | 201.5 | 201.9 | 202.5 | 203.7 |
| UD-IQ1_M | 192.62 GiB | 193.7 | 193.8 | 193.9 | 194.2 | 194.8 | 196.0 |
| UD-IQ1_S | 179.11 GiB | 180.2 | 180.2 | 180.4 | 180.7 | 181.3 | 182.5 |
| IQ2_S | 176.61 GiB | 177.7 | 177.7 | 177.9 | 178.2 | 178.8 | 180.0 |
| IQ2_XS | 175.47 GiB | 176.5 | 176.6 | 176.7 | 177.0 | 177.7 | 178.9 |
| UD-TQ1_0 | 158.79 GiB | 159.8 | 159.9 | 160.1 | 160.4 | 161.0 | 162.2 |
| IQ2_XXS | 152.79 GiB | 153.8 | 153.9 | 154.1 | 154.4 | 155.0 | 156.2 |
| IQ1_M | 137.32 GiB | 138.4 | 138.5 | 138.6 | 138.9 | 139.5 | 140.7 |
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.