WoolyAI Private Inference Stack is a new inference server for DGX Spark clusters, built to serve multi-agentic apps that run as a coordinated workflow across multiple models, modalities, tools, and context windows.
—Deploy dedicated
—Serve many agent types
—Serve multiple model families
—Run agents at low cost
Many agent types, models, and tools
Supervisor
Knowledge
Custom LLMs
Verifier
↓
Memory-aware multi-model serving across the cluster — not one static endpoint
↓
Private customer capacity · 4 Sparks
Spark 1
Spark 2
Spark 3
Spark 4
DGX Spark is smaller and more accessible than large data-center GPU systems, but large enough to support serious private inference workloads when paired with the right serving software.
Instead of deploying a separate expensive GPU stack for every model, InferenceStack for DGX Spark Cluster gives each customer a private inference environment that can serve many models of different sizes through one memory-model-residency-aware serving layer.
Benchmarks from internal tests on 2× NVIDIA DGX Spark systems — validated single-model throughput and multi-model agent workflow serving across a coordinated InferenceStack environment.
Large models can run with usable throughput on low-cost unified-memory GPU nodes.
Llama-Benchy • TP=2 • PP=2048 / TG=128 / depth=4096
Gemma 4 26B A4B C1 prefill on the two-node DGX Spark setup.
Nemotron 3 Nano Omni C4 aggregate decode across four simultaneous requests.
DeepSeek V4 Flash C1 decode on the two-node DGX Spark setup with no speculative decoding.
Switch models safely and keep decoding across an agent pipeline.
Simulated agent workflow • 1 endpoint • 3 models • C4 bursts
C4 decode throughput in the simulated workflow execution window.
C4 decode throughput after safe activation from DeepSeek to Gemma.
Safe-boundary activation from Gemma to Nemotron before the new model batch begins.
100% success, zero runtime failures across both ranks in the workflow run.
Internal validation • Metrics from July 2026 two-DGX-Spark Llama-Benchy and multi-model workflow reports.
Private enterprise multi-agentic inference: workloads that need privacy, multiple models, predictable cost, and dedicated capacity, but do not need a separate premium data-center GPU stack for every model in the workflow.