DeepSeek V4 Flash on DGX Spark

Measured / sourced results

What the numbers mean

What the numbers mean guidance: A public single-Spark result around 14 tok/s demonstrates feasibility, not a general speed guarantee. A practical planning band is low double digits for one node, with large variation from speculative decoding and expert kernels. Use one Spark only for a proven checkpoint recipe and context below 16K. Choose two nodes when the model is the product rather than an experiment.

What stands out

Single Spark: feasible only with a proven optimized checkpoint.

Recommended serving setup: two Sparks or at least 192 GB of fast accelerator memory.

Simpler alternative: a 27B–35B model when interactive latency matters.

Evidence quality / source

Evidence quality / source guidance: Model size comes from DeepSeek; the single-node speed anchor is a public source-reported configuration; the fit conclusion is engineering math. Do not multiply the 13B active-parameter count by bytes per weight to estimate resident memory; all experts normally remain stored.

What is unknown

• Do not multiply the 13B active-parameter count by bytes per weight to estimate resident memory; all experts normally remain stored.
• Checkpoint-specific compression and speculative decoding can materially change both fit and throughput.

Workload fit

Workload fit guidance: One Spark is for experimenters willing to tune memory and accept slower output. Teams that need stable APIs, long context, or multiple users should choose two Sparks or professional multi-GPU hardware. The model has 284B total parameters and 13B active. Even 4-bit-equivalent weights are roughly 142 GB before mixed-precision overhead, so a single 128 GB system needs special packing/offload rather than ordinary full residency.

Alternatives

Alternatives guidance: Qwen3.8-27B and Qwen3.6-35B are much easier, faster local models. Mistral Small 4 or gpt-oss-120b offers large-model capability that fits comfortably in 128 GB.

Planner CTA

Compare this recommendation against your model, context, concurrency, latency, and budget in the ComputeSage Planner.

Sources / methodology

Recommendations combine official specifications, public model/runtime documentation, adjacent benchmark observations, and clearly labeled engineering estimates. Review the ComputeSage evidence methodology.

• DeepSeek-V4-Flash model card
• NVIDIA DGX Spark specifications
• vLLM installation and hardware documentation
• llama.cpp multi-GPU documentation
• DeepSeek V4 Flash DGX Spark final benchmarks