Been putting MiniCPM5-2B to work on my DGX Spark. Built an app called Footnote that turns a research question into a living map of evidence. The setup: • Astra orchestrates through Codex CLI • MiniCPM5-2B workers read sources and extract quoted findings locally • Astra reviews the findings and connects them, with its reasoning visible First question: what’s keeping robots from being useful in ordinary homes? 21 sources. 31 claims. 11 unanswered questions. Click a claim and read the source passage. Click a connection and see why the AI made it. Live tokens/sec and concurrency are visible too. Then we added another paper. The map updated, and the original version stayed intact. All that extraction work from a 2-billion-parameter model. Next experiment: the AEON Qwen 27B model on my second Spark taking over orchestration. Already tested basic tool calls. A full local run is next. Building this one step at a time.