← AI Hashrate 中文

Can DGX Spark (GH200 480GB) run DeepSeek-V4-Pro?

NVIDIA · 480 GB LPDDR5X · 500 GB/s bandwidth · DeepSeek · 861.6B (active 49B) · MoE · model ctx up to 1000K

CC · WorkableComposite score: VRAM headroom × 0.5 + tok/s speed × 0.5. Q4 @ 8K context.
CUDA

Yes — the DGX Spark (GH200 480GB) (480 GB LPDDR5X) can run DeepSeek-V4-Pro. At Q4 with the default 8K context it needs ≈ 476.11 GB VRAM (weights 473.88 GB + KV cache 1.23 GB + 1 GB runtime overhead), which fits within 480 GB. Decode at Q4/8K: ≈ 6.5 tok/s (estimated, batch 1). FP16 needs ≈ 1725.42 GB — does not fit on this card. Estimates come from memory-bandwidth math; rows tagged measured override estimates. Relative ranking is more reliable than absolute tok/s. Methodology. Buy DGX Spark (GH200 480GB) on Amazon →

Fit & speed by quant and context

QuantContextVRAM neededFits?tok/s (decode)
Q44K475.49 GBYes6.5 ± est.
Q48K default476.11 GBYes6.5 ± est.
Q432K479.78 GBYes6.5 ± est.
FP164K1724.81 GBNo0.4 ± est.
FP168K default1725.42 GBNo0.4 ± est.
FP1632K1729.1 GBNo0.4 ± est.

Fits = weights + KV(ctx) + 1 GB overhead ≤ 95% of VRAM. Measured anchors are context-agnostic; the fit verdict is recomputed per context. A missing tok/s means the model is far beyond this card (offload-only territory).

VRAM breakdown at 8K context

QuantWeightsKV cacheOverheadTotal neededDGX Spark (GH200 480GB) VRAM
Q4473.88 GB1.23 GB1 GB476.11 GB480 GB
FP161723.2 GB1.23 GB1 GB1725.42 GB480 GB

Other GPUs that run DeepSeek-V4-Pro

No catalog GPU fully fits DeepSeek-V4-Pro at Q4/8K — see the model page for offload estimates. model page

All GPUs for DeepSeek-V4-Pro →

Other models for the DGX Spark (GH200 480GB)

All models on the DGX Spark (GH200 480GB) →