Google releases open MedGemma models for offline TB screening

At IDWeek 2026 in Washington, D.C., Google Research detailed its transition from proprietary medical imaging algorithms to open foundational models for tuberculosis screening. The team highlighted the deployment of these tools in six countries and their integration into Indonesia’s national health strategy to address data sovereignty and access gaps.

What happened

Tuberculosis remains a leading cause of death globally, claiming an estimated 1.23 million lives annually. A major barrier to detection is the scarcity of radiologists in high-burden settings, where imaging infrastructure may be days away from rural clinics or remote work sites. Over the past five years, Google Research and Google for Health have shifted focus from early benchmarks to deploying open foundation models that can be tuned and used at the point of care.

The journey began with proving that AI could match expert radiologists in reading chest radiographs. After publishing benchmark results in 2019 and seeing the World Health Organization recommend computer-aided detection in 2021, the team moved toward real-world validation. In September 2024, a prospective multi-site study published in NEJM AI confirmed that AI-assisted screening maintains high sensitivity across diverse populations while detecting other pulmonary abnormalities.

To bridge the gap between algorithm and clinical tool, Google partnered with Nexus AI, a Pretoria-based company. Nexus AI translated the chest X-ray algorithms into a certified diagnostic device, achieving CE MDR class IIb certification and ISO 13485 compliance. This tool is now active in approximately 40 healthcare facilities across South Africa, Sierra Leone, Côte d'Ivoire, the Democratic Republic of the Congo, Zambia, and Cameroon. Clinicians have used it to evaluate over 25,000 radiographs, reducing turnaround times from weeks to under 45 seconds.

How it works

The core technology relies on open-weight multimodal models that allow local health systems to build sovereign solutions. Google released MedSigLIP, a lightweight vision-language model that aligns medical images with clinical text, enabling developers to search or classify images using plain-language labels. Building on this, they launched MedGemma, a family of open-weight models with 4 billion and 27 billion parameter configurations.

MedGemma supports multimodal clinical reasoning, synthesizing radiographic findings with patient history to answer clinician queries in natural language. Crucially, these models run locally on-device without internet access. This capability allows health ministries to fine-tune the models on domestic datasets, keeping patient data within national borders. For example, Indonesia is using MedGemma to screen up to 50 million citizens annually, avoiding the data residency risks of proprietary cloud software.

Complementing the visual models, Google also developed Health Acoustic Representations (HeAR). Trained on over 300 million audio segments, HeAR enables non-invasive screening via cough sounds recorded on standard smartphones. This acoustic triage works alongside mobile X-ray screening to create an integrated diagnostic cascade for underserved populations.

Key details

• Global impact: Nexus AI’s tool has screened over 25,000 chest radiographs in six countries, with over 90 clinicians involved.
• Speed: The AI reduces diagnostic turnaround time from weeks to under 45 seconds per scan.
• Open models: MedGemma (4B and 27B parameters) and MedSigLIP are open-weight, allowing local fine-tuning and offline use.
• Adoption: MedGemma has been downloaded more than 10 million times since its launch.
• Sovereign deployment: Indonesia is fine-tuning these models on local infrastructure to screen 50 million citizens annually while maintaining data residency.
• Extreme environments: The technology operates four kilometers underground at the Mponeng Gold Mine in South Africa, enabling immediate triage for miners exposed to silica dust.

Why it matters

For engineers building health tech, this shift from closed APIs to open foundational models changes the architecture of deployment. It removes the dependency on constant connectivity and proprietary cloud services, which are often unreliable or legally problematic in global health contexts. By providing models that run on-device, Google enables local teams to create specialized solutions tailored to regional disease patterns without needing massive compute resources.

The emphasis on regulatory compliance and clinical validation also offers a roadmap for moving AI from research to production. Achieving CE MDR class IIb certification required rigorous audits and prospective clinical trials, demonstrating that safety and efficacy must be baked into the development lifecycle. This approach ensures that AI tools are not just accurate in benchmarks but trusted in bedside care.

Furthermore, the integration of multimodal reasoning allows for more nuanced clinical support. Instead of simple binary classification, models like MedGemma can interpret complex cases by combining image data with textual clinical history. This capability empowers healthcare workers who may not be specialist radiologists to make informed decisions, effectively democratizing expertise in resource-constrained settings.

What you can do

• Explore the open-weight MedGemma and MedSigLIP models through the Health AI Developer Foundations (HAI-DEF) resources.
• Evaluate offline inference capabilities for your own medical imaging projects to ensure functionality in low-connectivity areas.
• Study the regulatory pathway taken by Nexus AI, including ISO 13485 compliance and CE MDR certification, for your own medical device planning.
• Consider multimodal inputs, such as combining visual data with acoustic signals like cough sounds, to broaden diagnostic coverage.
• Engage with local health partners to understand specific workflow needs before designing AI interventions.
• Review the prospective validation studies in NEJM AI and The Lancet Digital Health to understand performance metrics in real-world cohorts.