Can DeepSeek V4 Pro run on the DGX Spark?
No. Even at 2-bit with 8K context, DeepSeek V4 Pro needs about 536 GB, more than the 109 GB usable on the DGX Spark.
Newer compressed-attention design: long-context figures are an upper bound.
Memory by quant and context
| Quant | 4K ctx | 8K ctx | 32K ctx | 128K ctx | Max tok/s (est., 8K) |
|---|---|---|---|---|---|
| 2-bit (Q2_K) • | 536 GB No | 536 GB No | 540 GB No | 553 GB No | — |
| 3-bit (Q3_K_M) | 729 GB No | 729 GB No | 733 GB No | 746 GB No | — |
| 4-bit (Q4_K_M) • | 890 GB No | 891 GB No | 894 GB No | 907 GB No | — |
| 5-bit (Q5_K_M) | 1062 GB No | 1063 GB No | 1066 GB No | 1079 GB No | — |
| 6-bit (Q6_K) | 1222 GB No | 1223 GB No | 1226 GB No | 1239 GB No | — |
| 8-bit (Q8_0) • | 1558 GB No | 1559 GB No | 1562 GB No | 1575 GB No | — |
| 16-bit (F16/BF16) | 2979 GB No | 2980 GB No | 2983 GB No | 2996 GB No | — |
• real GGUF file size from teamblobfish/DeepSeek-V4-Pro-GGUF. Speed is a bandwidth ceiling (273 GB/s), not a benchmark. Methodology.
More
DeepSeek V4 Pro on every GPU and Mac · Every model that runs on the DGX Spark · Other models that fit 64 GB VRAM · Try other settings in the calculator
Data: Hugging Face Hub (deepseek-ai/DeepSeek-V4-Pro), NVIDIA specs. Data as of 2026-11-17.