MMOOZZGG.com

Glossary

What is Google Gemma?

In short

Google Gemma is a family of open models from Google based on the research and technology behind Gemini; the fourth generation is released under the Apache 2.0 license and runs on your own server.

Models and products · Gemma

MMOOZZGG.com editorsChecked: 1 min read

For a child

Imagine a famous chef published a small recipe book that you can take home and cook from in your own kitchen, without going to the restaurant. The dishes are simpler, but the kitchen is yours, and nobody sees what you cook. Still, check the recipes: even a good book has typos.

In plain words

These are Google models that you can download and run yourself, not only reach through the cloud. The first generation came out on 21 February 2024 in two sizes, 2 and 7 billion parameters. The name comes from the Latin word for a precious stone. The fourth generation came out on 2 April 2026: it understands text and images, and three of its five models also understand audio. Example: a company puts the model on a server in its office and sorts incoming emails without sending them to someone else’s cloud.

For business

Companies in Russia need to tell Google’s cloud models from open weights. Gemma 4 came out under the Apache 2.0 license, which allows you to change the model and build your own products on it, and Google writes directly about running it locally without cloud infrastructure. The earlier generations, from the first to the third, including 3n, come with separate terms of use and a prohibited use policy; neither the license nor the terms contain a list of countries. The cost is the server and the graphics cards: compact models need little, and large ones need a lot of video memory. The hardware is chosen for the model size and the load.

For a specialist

A family of open-weight models from Google DeepMind and other Google teams. The first technical report, “Gemma: Open Models Based on Gemini Research and Technology” (2024), describes models with 2 and 7 billion parameters in pretrained and instruction-tuned versions. Gemma 4 has five sizes on four architectures: E2B and E4B with 2 and 4 billion effective parameters for phones and browsers, a multimodal 12B without a separate encoder, a 26B A4B mixture of experts where about 4 billion parameters work per token, and a dense 31B. The context window is 128,000 tokens for small models and 256,000 for medium ones, with reasoning modes, function calling and a system prompt. The weights come in BF16 and also in 8 and 4 bits, including with quantization-aware training. You can download them from Kaggle and Hugging Face.

Example

A legal team runs a fourth-generation model on its own server and tests how it pulls deadlines and amounts from contracts. The documents never leave the company’s network, and a lawyer checks a sample of the results.

Related terms

Questions and answers

Can you use Gemma for commercial purposes? The fourth generation is released under Apache 2.0, a license that allows you to change the model and build your own products on it. The first three generations have separate terms of use; read them before you deploy.
How is Gemma different from Gemini? Gemini is available only through Google services, while the open Gemma models can be downloaded and run on your own hardware. Google says they are built on the same research and technology as Gemini.

Sources

Not sure whether your company needs this?

AI solutions map
Contact us