What is Gemma?
The Harvard Chan AI Glossary
In our classes, we often pause to explain the ideas behind AI and the language used to describe it. These posts grew out of those conversations. They’re for anyone who would like to understand AI better, with no technical background assumed.
Gemma is Google’s family of AI models with downloadable weights. Weights are the numerical settings learned during training, the process through which a model learns patterns from examples. With appropriate software and hardware, downloadable weights allow a model to run outside a service operated directly by its developer.
Gemma is distinct from Gemini, although Google describes the families as drawing on related research and technology. Gemini names another model family and an assistant application. Gemma names models that can be incorporated into different applications and computing arrangements. Similar branding should not hide that difference.
One arrangement is to run a Gemma model locally through Goose. Local means that the model’s calculations run on the available computer. Goose provides the agent environment: the software organizing interaction, information and tools around the model. Speed and output depend on the particular model and machine.
Hardware matters because a computer must store the model’s numerical values and perform its calculations. Memory is the working storage available to those calculations. Different model sizes and numerical formats require different amounts. A model that is practical on one machine may be slow or unsuitable on another.
Local operation can give the user greater control over where model computation occurs, but it does not automatically make an entire application offline. A tool might search the web, contact a remote service or send logs elsewhere. The path taken by the information needs to be understood across the complete application.
Model releases also come with usage terms. Downloadable is a statement about availability, not a complete answer about licensing, redistribution or suitable uses. Check the specific model’s documentation rather than assuming that all releases have identical properties.
A healthcare teaching exercise could extend the local-model demonstration by asking a supported Gemma setup to summarize fictional referral notes, then checking omissions and response time on the available machine. That would examine one concrete arrangement without exposing patient records. It would not establish readiness for clinical use or prove that every part of the application stays offline. The useful lesson is how model choice, hardware and the surrounding tools jointly affect the work.
Source: Google — Gemma Models Overview.
If you would like to find out more about AI and improve your leadership skills, consider the AI in Health Care Certificate of Specialization from the Advanced Learning Academy at Harvard T.H. Chan School of Public Health. You can take an individual course or work toward the full certificate.
AI is a fast-moving field. This explanation reflects our class discussions and related material at the time of writing. Terminology, capabilities and tools may change.
For suggested additions or corrections, please email me at tpanch@hsph.harvard.edu.
Dr. Trishan Panch
Program Director, AI in Health Care
Harvard T.H. Chan School of Public Health — Advanced Learning Academy