We noticed something worth recording in the medical AI open-source community this week: OpenMed 3.0 shipped, while simultaneously posting 422 unresolved issues publicly seeking help. Its selling point isn't how strong the model is, but "never silently calling cloud APIs" — a growing anti-consensus judgment in the medical AI circle.
What this is
OpenMed is an Apache-2.0 licensed open-source medical AI toolkit — essentially a "local version of a medical AI stack." It runs on hospital servers or doctor workstations, data never leaves premise, and it absolutely does not call cloud vendors like OpenAI or Anthropic.
Version 3.0's centerpiece is "patient journey": patient data scattered across FHIR (medical data exchange standard), HL7 (hospital-to-hospital communication protocol), lab CSVs, and radiology reports is stitched into one timeline. The key design principle is "refuse to guess" — even a 99% identity match does not auto-merge, human review is required; when two sources conflict, both are kept with conflict labels. In the demo, 5 synthetic cases in chaotic formats ran end-to-end in 0.07 seconds.
Industry view
Supporters say this path is right. The "silent cloud fallback" of medical data is a dangerous decision nobody signed off on; OpenMed welds that door shut, which is a direct win for regulators and hospital CIOs (Chief Information Officers).
Opposition is sharp. The entire project currently runs mainly on 1 maintainer (maziyarpanahi), with 422 issues waiting for someone to pick up — a dangerous signal. Once the core maintainer's attention shifts, this "data never leaves the hospital" moat collapses. Version 3.0's built-in summarization feature is explicitly labeled "synthetic data only, not for real patients" — far from clinical deployment. More critically, mainstream medical AI investment (OpenAI, Alibaba, Baidu) all bet on the cloud, treating on-premise as "politically correct but commercially unsustainable" — OpenMed is more community idealism than industry future.
Impact on regular people
For enterprise IT: The local AI path has been proven technically feasible, but only on the condition of having an internal maintenance team — otherwise the risk of an open-source project going stale gets amplified.
For individual careers: Engineers who know FHIR and HL7 will see new demand scenarios — the "vertical industry + local AI" combination is almost a blank slate in China right now.
For consumer markets: Ordinary patients won't use OpenMed in the short term, but the direction it represents — "my medical records shouldn't be quietly uploaded" — will become a new privacy baseline, in turn forcing commercial medical AI products to add local options.