They arrived on a Saturday. Two NVIDIA DGX Spark units, picked up from the post office after sitting in customs for days. I'd been planning the setup for weeks — models pre-downloaded, Docker images queued, architecture docs written. I was ready.
What I wasn't ready for was this:
The unboxing was satisfying. The hardware is beautiful — compact, dimpled metal chassis, surprisingly heavy for its size. The kind of thing that makes you feel like you're living in the future.
And then I looked at the plug.
Shipped from the UK, apparently — universal voltage (100–240V), but the wrong end for a US outlet. It was Saturday evening. Hardware stores were closed. I ordered NEMA 5-15 to IEC C5 cables from B&H Photo overnight, and the Sparks sat on the desk for two more days looking at me.
They powered on Monday morning.
What's running on them
Spark 1 is the always-on agent brain — Nemotron-3-Super-120B-A12B (NVFP4, ~80GB) via TRT-LLM, feeding the MetaClaw routing layer that decides what goes local vs. what escalates to Claude. Spark 2 handles async response scoring and eventually RL fine-tuning. 256GB pooled via NVLink-C2C for the big runs.
First real workload: OpenViking memory indexing. 713 items, 1.27M VLM tokens. The Sparks earned their keep on day one.
GTC keynote is Monday. Feels like good timing.
The plan
- Week 1: Nemotron live on Spark 1, targeting 40–50% local query routing
- Week 2: Response judge on Spark 2 — async quality scoring, data collection only
- Week 4+: RL training pipeline, gated behind explicit approval
- Long-term: 88.7% local routing (the OpenJarvis benchmark for "what can stay on-device")
- Phase X: PuppyPi robot camera feeds → Spark 2 vision inference
Two DGX Sparks. One UK plug problem. Fully resolved.
More soon.