What is Gemma AI? Google’s open Gemma 4 models explained

Gemma AI is Google DeepMind’s open-weight model family, and Gemma 4 is the version to learn now. You download the weights free and run them yourself. This guide covers the five sizes, the licence, and how to run one.
What is Gemma AI?
Gemma AI is Google DeepMind’s family of open-weight models that you download and run on your own hardware, from a phone to a server. It is built from the same research behind Gemini, and the current generation, Gemma 4, is free to use commercially.

The name comes from the Latin word gemma. The first models arrived in February 2024 as small 2B and 7B text models. It works like any large language model, except the model lives on your device. Weights are free on Hugging Face and Kaggle, and a Gemma overview traces the family from the start.
What are the Gemma 4 model sizes?
Gemma 4 ships in five sizes, E2B, E4B, 12B Unified, 26B A4B, and 31B, so you pick the model your hardware can hold. The two small sizes run on phones and laptops, while the three larger ones suit workstations and GPUs with longer context.

Google’s release log lists the first four sizes on 31 March, announced 2 April, with 12B Unified arriving in June. The Gemma 4 model card confirms the split: 128K tokens on E2B and E4B, 256K on the larger three. All five take text and image, the small three also take audio, in over 140 languages.
| Size | Parameters | Context | Best for |
|---|---|---|---|
| E2B | 2B effective | 128K | Phones, edge |
| E4B | 4B effective | 128K | Laptops, apps |
| 12B Unified | 12B | 256K | All-round work |
| 26B A4B | 26B, 4B active | 256K | Fast server inference |
| 31B | 31B dense | 256K | Best quality, one GPU |
Why does the Apache 2.0 licence matter?
Gemma 4 is the first Gemma generation released under the Apache 2.0 licence, which lets you use, change, and sell products built on it. Earlier generations stay under Google’s own Gemma Terms of Use, so the generation you pick decides what your lawyers will say.

Apache 2.0 is a permissive licence: commercial use is allowed, the weights can change, and there is no royalty. Gemma 3 and 3n sit under the Gemma Terms of Use and a prohibited-use policy. Every generation carries the same name, so check the licence line first.
| Generation | Licence | What it means |
|---|---|---|
| Gemma 4 | Apache 2.0 | Use, modify, sell freely |
| Gemma 3 | Gemma Terms of Use | Custom terms, read first |
| Gemma 3n | Gemma Terms of Use | Custom terms, read first |
How do you run Gemma AI on your own device?
You run Gemma AI by downloading the weights from Hugging Face or Kaggle, then loading them in a runner such as Ollama or LM Studio. The small sizes need a phone or laptop, while the big ones want a graphics card with more memory.

Start with the instruction-tuned version, marked it. Ollama lists Gemma 4 tags such as gemma4:26b, and LM Studio offers point-and-click setup. Gemma 3 repos on Hugging Face are gated: accept the terms first. Once downloaded, the model runs offline, so nothing you type leaves the machine. Fine-tune it with Unsloth or Colab.
| Size | Where it runs | How to run it |
|---|---|---|
| E2B | Phone, 6 GB RAM | Ollama, LiteRT |
| E4B | Laptop or phone | Ollama, LM Studio |
| 12B Unified | Workstation GPU | Ollama, vLLM |
| 26B A4B | 24 GB GPU | vLLM, Ollama |
| 31B | High-end GPU | vLLM, Ollama |
Gemma AI vs Gemini: what is the difference?
Gemma AI is the open-weight model you run yourself, while Gemini is Google’s closed, hosted frontier service. They share research lineage, but Gemma gives you the weights and Gemini gives you an API, and that changes cost, privacy, and control.

Gemini is a hosted service you pay to use, while Gemma runs on your machine, so prompts never leave it. Gemini leads on power, Gemma wins on local data. Gemma is not Gemini 4, a closed product. Like every model, Gemma can hallucinate, so check facts before you act on them. It is a tool, not general intelligence.
| Point | Gemma AI | Gemini |
|---|---|---|
| Access | Download weights | API or chatbot |
| Cost | Free weights | Pay per use |
| Privacy | Offline | Prompts go to Google |
| Top capability | Strong for its size | Frontier models |
| Customisation | Fine-tune weights yourself | Limited tuning |
What can you build with Gemma AI?
You can build offline chatbots, coding helpers, document question answering, and fine-tuned support bots with Gemma AI. The small sizes suit apps that must work without the internet, while the large sizes suit servers that need long context and stronger reasoning.
Start with document question answering: index your manuals and let Gemma answer from them, the idea behind retrieval-augmented generation. DiffusionGemma, an experimental model on the Gemma 4 26B backbone, writes whole blocks of text in parallel. Google’s DiffusionGemma guide reports up to four times faster generation.
What are the honest limits of Gemma AI?
Gemma AI has old knowledge cutoffs, a tiny model that struggles with hard reasoning, and a licence that changes by generation. It also ships as weights only, so hosting, safety filters, and updates are your job to handle, not Google’s.
Gemma 3 data stops in August 2024, and 3n in June 2024, so newer facts need retrieval. The 270M model is tiny and weak, so keep it to simple jobs. Open weight is not open source, and Gemma 3 and 3n need a terms check before you ship. DiffusionGemma trades quality for speed.
| Limit | What it means | Workaround |
|---|---|---|
| Old cutoff | Gemma 3 stops Aug 2024 | Add retrieval |
| 270M model is tiny | Weak reasoning | Simple jobs only |
| Licence split | Gemma 3/3n custom terms | Read the model card |
| DiffusionGemma is experimental | Lower quality | Use Gemma 4 |
Frequently asked questions about Gemma AI
The questions below cover Gemma AI pricing, how it differs from Gemini, running it on a phone, commercial use, picking a size, and offline use. Each answer gives the short version first, so you can scan the list and read only the detail you need.

Is Gemma AI free?
Yes. The weights are free from Hugging Face and Kaggle. You pay only for hardware and electricity.
Is Gemma AI the same as Gemini?
No. Gemma is the open-weight family you run yourself, while Gemini is Google’s closed, hosted service.
Can I run Gemma AI on my phone?
Yes, with E2B, built for phones and edge devices, or E4B on a stronger phone. Use Ollama.
Can I use Gemma AI for commercial products?
Gemma 4, yes, under Apache 2.0. Gemma 3 and 3n fall under the Gemma Terms of Use, so read those first.
Which Gemma model should I pick?
Pick E2B for a phone, E4B for a laptop, and 12B, 26B A4B, or 31B for a server.
Does Gemma AI need the internet?
No, not once the weights are downloaded. The model runs fully offline, though live retrieval still needs a connection.
Try Gemma AI: download the smallest size for your device and use it on real work. If it answers well offline, it costs nothing to run. For site owners, AI answers are the front door to content, so generative engine optimization matters too.