Enthusiast builds 16x DGX Spark cluster to run frontier open models

A Reddit user known as ciprianveg is building what could be one of the most powerful personal AI clusters to date: a 16-node setup based on NVIDIA's DGX Spark hardware, specifically the ASUS GX10 variant. Each node packs the GB10 Grace Blackwell superchip, designed for local AI inference and fine-tuning. The cluster is networked via a MikroTik CRS804-4DDQ switch, using four breakout cables that convert 400Gbps uplinks to 100Gbps per node, giving the user a high-bandwidth, low-latency fabric for distributed inference. Their stated goal is to run frontier-level open-weight models locally—including DeepSeek V4 Pro, Kimi K3, and upcoming releases like GLM 5.5 and MiniMax M4—without relying on cloud APIs.

The user's plan is to split the cluster into two 8-node configurations, allowing two separate large models to run concurrently. However, the hardware also leaves room for ambitious workloads: with all 16 nodes combined, they aim to handle models with over 2 trillion parameters. That's a scale normally reserved for data centers, enabled by DGX Spark's 128GB unified memory per node and modern interconnect options. While the setup is enthusiast-grade and not production-tested, it signals a broader trend: as open-weight models grow and hardware costs stabilize, serious hobbyists can now experiment with agentic AI and massive LLMs in their own homes—pushing the boundaries of what's possible outside the cloud.