Gemma

advancedwidespread

Gemma is Google's family of open-weight AI models available for free download, which can run even on your own laptop or phone without an internet connection.

In plain terms

Gemma is the Google DeepMind model family, the younger sibling of Gemini. You access Gemini on Google's own servers, while Gemma can be downloaded by anyone and run on their own machine.

It is like the difference between dining at a restaurant and taking a recipe home. Gemini is the restaurant: convenient and powerful, but you cannot see into the kitchen, and every order goes through the company. Gemma is the recipe and ingredients: you cook at home, tailor it to your taste, and no one sees what you eat.

The Gemma family exists in several sizes: the smallest can run on a phone, while the largest require powerful graphics cards. There are also specialized versions, for example for medical texts and images (MedGemma), programming (CodeGemma), or content filtering (ShieldGemma).

This is important because it allows a small company, hospital, or office to use a language model without sensitive data leaving the building.

How it works

Google trains the model, building on research results and technologies also used in the development of Gemini, but on a much smaller scale.
The finished model's 'weights' are published, for example on Hugging Face and Kaggle. Weights are numbers formed during training, ranging from hundreds of millions to many billions depending on the size variant.
You download a size suitable for your machine and launch it with a runner (for example Ollama or LM Studio). From then on, responses are generated on your own machine.
If needed, you can further train it with your own examples (this is fine-tuning) so it better understands your field or your company's style.
Developers can embed it into an application or an agent: newer versions understand images and can call external tools.

An everyday example

An accounting firm does not want to send clients' invoices to a foreign cloud. Their IT specialist therefore installs Gemma on an office computer and uses it to generate summaries from the documents. The data thus remains entirely on their own machine.

Who is affected

developerYou get a model that is free to download and customizable, which you can integrate into your own application or agent. You do not pay usage-based fees; only the cost of the machine required to run it is charged. → Code

entrepreneur, executiveYou can introduce AI so that trade secrets and customer data remain in-house and you do not depend on a single provider's prices.

patient, doctorWith healthcare variants, hospitals can develop AI tools on their own machines, so patient data does not go to external servers. → Health

parent, student, teacherYou can try and learn how a language model works on an average computer without a subscription. → Education & Family

A common misconception

MythA Gemma ugyanaz, mint a Gemini, csak ingyen.

In realityGemma builds on related technology but is a much smaller and less capable model. In return, it runs on your own machine and is freely adaptable.

Limits and risks

Smaller variants make mistakes and hallucinate facts more often than large cloud models, so always verify important answers.
It understands Hungarian, but smaller models formulate Hungarian noticeably worse than English.
Larger variants require a powerful graphics card and lots of memory, while installation and maintenance are your responsibility.
"Open" does not mean unlimited freedom: license terms vary by generation; read them before business use.

Similar, but not the same

Gemini – Gemini is Google's large, closed model family, accessible on the company's servers. Gemma, however, is downloadable and runs locally.
Open-source software – With Gemma, the final model weights are public, but the training data and full training process are not; hence it is more accurately called an 'open-weights' model.
Llama – Llama is Meta's downloadable model family with similar goals, meaning a competitor, not the same product.

Related concepts

Broader

open-weight model, Large language model, Foundation model, Kis nyelvi modell

Narrower

MedGemma, CodeGemma, PaliGemma, ShieldGemma, EmbeddingGemma

On the edge

Gemini, Llama, Google DeepMind, Local Execution, Fine-tuning

A brief history

Google introduced the first Gemma models in February 2024. The name means gemstone in Latin. Gemma 2 was released in mid-2024. In 2025, Gemma 3 could already understand images and handle more than 140 languages, while Gemma 3n was designed specifically for phones. Gemma 4 was released in 2026.

Where things stand

By the end of 2026, Gemma is one of the most downloaded open-weight model families, with tens of thousands of community variants. It is used mainly by developers, researchers, and organizations sensitive to data protection. The average user encounters it embedded in applications.

Status: 2026 November

Also known as

Gemma · Google Gemma · Gemma 3 · Gemma 3n · Gemma 4 · Gemma-modell · Gemma model family · Gemma open models

Sources

This entry was machine-edited and has undergone machine expert review; the “In the news” section updates automatically from Agent Age stories.