LUMEN is an open-source neonatal danger-sign screener published on the Alibaba Cloud Project Hub on 29 October and built by an independent developer. It follows the WHO IMNCI protocol and returns 21 danger-sign checks, a red, yellow or green classification, a dosing and referral plan, and a follow-up date. A visit runs as one guided flow: voice intake in 12 languages, seven camera-guided photographs, a tap-to-count breathing rate that flags 60 or more breaths per minute, and a 10-second cry recording. Five agents handle vision, audio, classification, triage and memory. Red and yellow cases are queued for a supervisor to sign off. Requests are routed by a failover policy that weighs connectivity, latency and privacy: online they go to qwen3-vl-plus, qwen3-omni-flash and qwen3.7-plus through Model Studio; offline they go to LUMEN-Q, a Qwen3-VL-4B-Instruct fine-tuned in two QLoRA stages and exported to 4-bit GGUF, running through llama.cpp on a laptop and on Android through a native bridge. The five WHO clinical calculations are exposed as an MCP server, so the model selects a tool instead of doing arithmetic. On 150 held-out neonatal images, parseable clinical answers rose from 65 to 100 percent and triage accuracy roughly quadrupled, according to the project write-up.
Our takeParseable answers and correct answers are different measurements, and the write-up reports the first prominently: a model that always emits a well-formed label can reach 100 percent on format while staying wrong. The 150-image set is also small and drawn from one collection. What is genuinely reusable is the split: deterministic clinical arithmetic behind tool calls, and one streaming core shared by cloud and device so the router cannot change the logic. Prospective field validation is the step that would settle accuracy.