← DGX Spark hub · Single · Dual · Quad

Quad DGX Spark (~512GB) for local RAG

Limited evidence Four-Spark setups are rare. The results below come from a handful of owners, with very different configurations.

What fits

Model	Quant / engine	Measured decode	Source
MiniMax-M3	AWQ-INT4 (~241GB), vLLM TP=4 + EAGLE3	~38 tok/s single stream; 79.4 aggregate at 4 concurrent	tonyd2wild
DeepSeek-V4-Flash	vLLM TP=4 over a switch	~90 tok/s single stream, ~2,500 tok/s prefill	NVIDIA forum
Qwen3.8-Flash-Next	NVFP4, TP=4 + expert parallel	40.5 median / 54.2 peak tok/s; 262 aggregate	NVIDIA forum
MiniMax M2.5	vLLM TP=4 (NVIDIA example)	No published benchmark yet	NVIDIA vLLM playbook
Llama 3.1 405B	4-bit	No published benchmark yet	—

How to link the units

NVIDIA's Connect multiple Sparks through a switch playbook covers this setup, and the NCCL playbook tests two, three, or four Sparks. NVIDIA's vLLM cluster playbook supports two to four Sparks and says a switch is required for four. A three-unit ring needs no switch (playbook).
Owners use 200G switches such as the MikroTik CRS812 with breakout cables. Measured overhead was about 0.7% vs direct cables (source).
A switchless 4-node ring cannot run stock NCCL TP=4, because non-neighbor ranks have no direct link. Patched NCCL stacks work around this (source, source).

Recommended local RAG stack

Inference: vLLM TP=4 for one very large MoE model. Alternatively, two independent TP=2 pairs, which one owner found gave more total throughput (source).
Retrieval: embeddings, reranker, Docling, and the vector DB as on dual.
Front-end: Dify or Open WebUI.

Models from our top 10 that fit

GLM-5.3-Flash: Yes, near-full precision fits (inferred) unverified
Qwen3.8-27B: Yes
DeepSeek V4.1 Flash: Only if quantized, untested unverified
MiniMax M3.1 Flash: Likely at 4-bit (unverified) unverified
Qwen3.6-35B-A3B: Yes
Nemotron 3.5 Super VL: Yes
gpt-oss-120b: Yes
Gemma 4 26B-A4B: Yes
MiMo-V2.6-Flash: Yes unverified

Spark fit is estimated from published memory sizes, not tested on Spark hardware. See Best local models.

Best for

Experiments with the largest open MoE models, long-context serving with a shared prefix cache, and labs. For most document Q&A, one or two units do the job at far lower cost.

Gotchas

Single-stream speed from 2→4 nodes rises only a few percent to roughly 1.5×, depending on the model (source).
Keep each node's disk below 95%, or Ray kills the engine (source).
Driver and NCCL versions matter. One installer pins driver 580 because of issues reported on 590+ (source).

Benchmarks are quoted from the linked sources with their settings; we have not re-run them. Where no result exists we say "no published benchmark yet". Last updated 2026-10-24.