Taking on new work
Lab & internal · 2026

Jetson edge inference nodeLive

An NVIDIA Jetson Orin Nano running local language models as the lab's free, private default brain for classification and first-pass triage.

Cloud models are fast and sharp; they also cost money and see your data. The Jetson is the answer for everything that does not need a frontier model: news bucketing, first-pass classification, and anything that should never leave the network.

Notes from running small models for real

  • A 3B model on the default 4K context silently returned empty results when the prompt overflowed. Raising the context window and shrinking the batch fixed it; the lesson was to measure output, not trust it.
  • Thermal and fan telemetry from the board is read over SSH and shown on the status board, because a throttled node looks exactly like a slow model.

At a glance

StatusLive
TypeLab & internal
Year2026

AI

OllamaQwen 2.5Local inference

Stack

NVIDIA Jetson Orin NanoUbuntuOllamaPrivate mesh
Related work