On April 2, 2026, Google DeepMind launched Gemma 4 — a family of open-source artificial intelligence models that anyone can download, install and use completely free of charge, without a subscription, without a paid API and without any data leaving your device. It is Google's most significant move in the open-source AI space since it started releasing Gemma models in 2024, and the first under an Apache 2.0 license — the most permissive on the market — meaning it can also be used in commercial projects without restrictions.
With more than 400 million cumulative downloads across the entire Gemma family since its launch, Gemma 4 arrives in four variants of different sizes designed to run on hardware as varied as a mid-range smartphone, a laptop, a gaming PC with a dedicated GPU or an enterprise server.
What Gemma 4 is and how it differs from Gemini
Gemma 4 and Gemini are related but distinct technologies. Gemini is Google's AI assistant used from a browser or app — the equivalent of ChatGPT, which runs on Google's servers and is accessed via the internet.
Gemma 4 is the model you can download and run yourself, on your own device, without sending data to any server. It is built with the same research and technology that powers Gemini 3, but in more compact versions optimized to run locally.
The advantage is clear: complete privacy, no cost and no need for internet after the initial download.
The four Gemma 4 models
Google launched four variants with different sizes and use cases:
Gemma 4 E2B — the smallest, with around 2 billion parameters. Designed to run on smartphones, Raspberry Pi and IoT devices with near-zero latency. Natively processes text, images and audio. 128,000 token context window. This is the model to install on your phone.
Gemma 4 E4B — a slightly larger version with around 4 billion effective parameters. Same profile as the E2B but with greater reasoning capacity. Also suitable for smartphones and basic laptops.
Gemma 4 26B A4B — a Mixture of Experts (MoE) architecture model: it has 26 billion total parameters but only activates around 4 billion during each inference. This makes it much more efficient than its size suggests. Recommended for gaming PCs and laptops with a GPU. 256,000 token context window.
Gemma 4 31B — the most powerful model in the family, with 31 billion parameters in a dense architecture. It ranks third worldwide in the Arena AI open model leaderboard. For workstations with a dedicated GPU.
All models process text and images. The E2B and E4B models also natively process audio.
How to install Gemma 4 on your PC — step by step
The easiest way for users without technical experience is to use Ollama, a free tool that downloads and runs AI models locally with a single command.
On Windows, Mac or Linux:
1. Go to ollama.com and download the installer for your operating system. It's free.
2. Install it like any regular program.
3. Open the terminal (on Windows: search for "cmd" or "PowerShell") and type:
`ollama run gemma4:e4b`
for the 4B model, or `ollama run gemma4:2b` for the lightest version.
4. Ollama automatically downloads the model (between 2 GB and 4 GB depending on size) and opens a direct chat in the terminal.
You can also download it directly from Hugging Face (huggingface.co) by searching "google/gemma-4" or from Kaggle if you prefer a notebook environment.
What hardware you need on PC:
- Gemma 4 E2B / E4B: any modern PC with 8 GB of RAM. Runs on CPU without a GPU, though more slowly.
- Gemma 4 26B: GPU with at least 16 GB of VRAM (RTX 4080, RTX 5080, or Mac with 24 GB unified memory).
- Gemma 4 31B: 24 GB of VRAM or more (RTX 3090, 4090, 5090).
If you don't have a GPU but want to try the larger models, you can use them online for free on Google AI Studio (aistudio.google.com) without installing anything.
How to install Gemma 4 on your phone
The E2B and E4B models were specifically designed to run on smartphones. Google also simultaneously launched Gemini Nano 4, the optimized version for mobile devices that some manufacturers will integrate directly into the operating system.
To install Gemma 4 on Android today:
Option 1 — Third-party apps: apps on Google Play like AI Runner or MLC Chat allow you to download and run Gemma models locally on the device. Search "local LLM" on Google Play and you'll find several compatible options.
Option 2 — Termux: if you have basic terminal knowledge, Termux (free app on F-Droid) lets you install Ollama on Android and run Gemma 4 E2B directly.
What phone you need:
- Gemma 4 E2B works on any modern Android with at least 6 GB of RAM — which covers the vast majority of mid-range and high-end phones from 2023 onwards.
- The model is around 2 GB of storage to download.
On iPhone, the situation is more limited due to iOS restrictions on running external code, though specific apps like LM Studio are working on iOS support.
What Gemma 4 can do
Once installed, Gemma 4 can:
Answer questions, write text, summarize documents and translate between more than 140 languages — including full Spanish support. Analyze images: describe photos, read text in images, identify objects. The E2B and E4B models also transcribe audio. Generate and review code in any programming language. Function as an autonomous agent: with native function calling, it can connect to external tools, APIs and databases to execute multi-step tasks without constant supervision.
All of this without an internet connection, without data leaving your device and at zero ongoing cost after the initial download.
Why this matters
The Apache 2.0 license is the most significant change in Gemma 4 compared to previous versions. Gemma 1, 2 and 3 had custom licenses that prohibited or restricted commercial use, causing corporate legal teams to reject them for production use. Apache 2.0 eliminates those restrictions: any company or developer can use Gemma 4 in commercial products, modify it, redistribute it and fine-tune it without paying royalties or asking permission.
Combined with its ability to run locally — including on mid-range smartphones — Gemma 4 is the most accessible option for anyone who wants AI running on their own device, without depending on any cloud service and without paying a monthly subscription.
Comments
💬 Log in to comment💬 Join the conversation and log in to comment.