Open models go to work: clinics, hospitals, and enterprises choose auditability

The clinical test

Google Research describes a tuberculosis screening effort that has moved from research tools to prospective clinical trials and regulatory clearance with local partners. Tuberculosis claims an estimated 1.23 million lives each year, according to Google Research. The team released MedGemma, a family of open-weight multimodal models in 4B and 27B parameter configurations, and says it has seen more than 10 million downloads worldwide since launch. Downloads are an adoption signal, not a clinical outcome, but the scale matters: an open-weight medical model that developers can actually run is a different object from a demo that only its maker can operate. m:…6314

The deployment detail is carried by a partner, Nexus AI. Google Research says Nexus AI achieved CE MDR Class IIb certification under EU medical device regulation and certified compliance with ISO 13485 for medical device quality management. Its tool screens chest X-rays for lung diseases in as little as 45 seconds, and operates in approximately 40 healthcare facilities across six countries. The Indonesian Ministry of Health, Google Research says, is fine-tuning MedGemma and MedSigLIP on domestic chest radiographs to support its national tuberculosis initiative. That is the open-weight thesis in practice: a shared model, locally fine-tuned, under local regulatory clearance. m:…6314

Google Research also describes HeAR, or Health Acoustic Representations, trained on over 300 million audio segments to enable non-invasive acoustic screening from cough sounds recorded on a standard smartphone microphone. The combination — open-weight vision and language models, acoustic screening, and local fine-tuning — suggests a pattern for regulated AI deployment. The model is open, the deployment is local, and the certification is local too. A closed model that cannot be fine-tuned on domestic data or inspected by a local regulator fits that pattern poorly. m:…6314

The hospital argument

Ambience Healthcare co-founder Mike Ng makes the general version of the same argument for hospitals. AI capabilities are advancing rapidly, he argues, but bringing those advances into everyday patient care remains a challenge. Health systems are introducing AI across documentation, clinical decision support, coding, and care coordination, according to Ambience. Each new application, Ng argues, faces the same underlying challenges: accessing reliable clinical context, working across existing systems, and operating safely at scale. Without shared infrastructure, health systems risk rebuilding those foundations for every new application. m:…8647

Ambience's answer is Chorus, its clinical AI infrastructure, designed around three capabilities: a System of Context, a System of Action, and a System of Governance. The vision, in Ng's account, is an open foundation that health systems, researchers, and developers can build on, allowing healthcare organizations to adopt new AI capabilities without starting from scratch each time. Ng's argument draws on six years of building alongside leading health systems, according to Ambience. That is a vendor argument, and it should be read as one — Ambience is selling the foundation it describes. But the problem it names is independently visible in the Google Research story: without shared context and governance layers, every deployment is a bespoke project. m:…8647

The enterprise echo

The enterprise version of the same preference surfaced on All-In, where Palo Alto Networks CEO Nikesh Arora joined the hosts to discuss AI spending. Enterprises are gravitating toward open-source and open-weight AI models to ensure explainability, auditability, and verifiability, according to the discussion. The spending context the hosts gave was concentrated: 1% of companies represent 90% of all AI spending, and 80% of that top 90% comes from unprofitable venture-backed companies. Whether those concentration figures hold exactly matters less than the direction they point: if most spending is concentrated and much of it is unprofitable, enterprise buyers have a reason to prefer models they can inspect, audit, and run themselves. m:…6532

The same discussion noted that the PHP2Shell exploit was identified and solved using persistent AI compute, and that AI models are increasingly capable of chaining vulnerabilities to find attack paths in enterprise infrastructure. That is a security argument for auditability as well. A model whose behavior cannot be explained or verified is harder to defend when it is also being used to attack. The open-weight preference, in this telling, is not ideological. It is operational: enterprises and health systems want to see how the model works, prove what it did, and run it where their data lives. m:…6532

What open wins, and what it does not

Taken together, the three sources describe open weight as a deployment strategy, not a research preference. Google Research shows an open model reaching tens of facilities and millions of downloads. Ambience argues hospitals need a shared foundation rather than a new bespoke system per application. All-In's guests say enterprises want models they can explain and verify. What open does not solve, on this evidence, is the hard part of any regulated deployment: local validation, certification, and governance. The open model makes those steps possible. It does not do them. The next test for open-weight medical and enterprise AI is not downloads, but how many deployments clear local regulatory and audit bars — and how many of those stay open once they do. m:…6314 m:…8647 m:…6532