Gemma 4: What It Is, the Sizes, the Licence and Where to Get It

Google calls them its most intelligent open models to date, built from the same research and technology as Gemini 3. One detail trips people up: the launch post named four sizes, and the DeepMind page now lists five, because a 12B has been added since launch.

Gemma or Gemini: the difference in plain words

Gemini is Google's model you use through Google's own services. Gemma is open weights: you download the model files and run them where you like, on a phone, a laptop, your own servers or a cloud account. Google says the licence gives you full control over your data, infrastructure and models, on premises or in the cloud.

That's the reason a business would look at Gemma at all. If your data can't leave your building, or you want a model that works offline, a model you host yourself is the route, and Gemma 4 is Google's version of it.

The five sizes and what each one runs on

Size	Built for	What to know
E2B and E4B	Phones, Raspberry Pi, NVIDIA Jetson Orin Nano	Effective 2 and 4 billion parameter footprint; run offline; take audio input; 128K token context
12B	Personal computers	Listed on the DeepMind page; not in the launch post
26B Mixture of Experts	Personal computers	Activates only 3.8 billion parameters per response, so it's faster; up to 256K tokens
31B Dense	Personal computers and servers	Highest quality and the base for fine-tuning; unquantized weights fit one 80GB NVIDIA H100

The E2B and E4B models were built with Google's Pixel team and with Qualcomm Technologies and MediaTek, to save memory and battery and to respond with near-zero latency on the device. At the other end, quantized versions of the bigger models run on consumer GPUs, so you don't need data-centre hardware to try them.

At launch, Google said the 31B ranked third among open models on the Arena AI text leaderboard and the 26B ranked sixth. Leaderboards move fast, so treat that as a launch snapshot.

What the models can do

• Multi-step reasoning, with native function calling, structured JSON output and native system instructions for building agents.
• Offline code generation.
• Video and image understanding, which covers reading text in images and reading charts.
• Speech recognition from audio input on the E2B and E4B models.
• Training across more than 140 languages.
• Long context: 128K tokens on the edge models and up to 256K on the larger ones, enough to pass a long document or a whole code repository in one prompt.

For Android teams, Google says developers can prototype agent flows in the AICore Developer Preview for forward compatibility with Gemini Nano 4, and use Gemma 4 for Agent Mode in Android Studio.

The Apache 2.0 licence and why it changed

Gemma 4 is released under Apache 2.0, which allows commercial use. Google says the switch came from developer feedback about restrictive terms on earlier releases. It also says Gemma 4 goes through the same infrastructure security protocols as its proprietary models.

Gemma's community is large. Google says earlier Gemma models were downloaded over 400 million times, producing more than 100,000 variants it calls the Gemmaverse. Two fine-tuning examples it gives are INSAIT's Bulgarian-first language model, BgGPT, and Yale University's Cell2Sentence-Scale work on cancer therapy pathways.

Where to try Gemma 4 and download the weights

You can try it before you download anything. Google AI Studio hosts the 31B and 26B models, and Google AI Edge Gallery hosts E4B and E2B. The DeepMind page lists Hugging Face, Ollama, Kaggle, LM Studio and Docker as places to get the weights, and Keras, Google AI Edge, Cactus, Google Kubernetes Engine and Unsloth among the ways to run, train and deploy it.

If you want to compare it with other models you can run yourself, my guide to where to download free AI models lays out the options. Ollama is one of the listed download routes, and my Ollama walkthrough covers the set-up.

Day-one support named in the launch post includes Hugging Face Transformers, vLLM, llama.cpp, MLX, Ollama, NVIDIA NIM and NeMo, LM Studio, Unsloth, SGLang, Docker and Keras, a long list of the tools developers already work with.

Hardware, fine-tuning and production

Google says Gemma 4 is optimised for NVIDIA hardware from Jetson Orin Nano up to Blackwell GPUs, for AMD GPUs through ROCm, and for Google's Trillium and Ironwood TPUs. To customise it, Google points to training and fine-tuning in Google Colab, Vertex AI or even a gaming GPU. For production it points to Vertex AI, Cloud Run, GKE, Sovereign Cloud and TPU-accelerated serving.

Odd places Gemma 4 has turned up

Two recent Google posts show how far the small models reach. In Project Suncatcher, announced on 30 October 2026, a prototype satellite ran an onboard Gemma 4 model that answered a prompt asking for a haiku about being in space. And Google says it used Gemini 4 Argon to speed up Gemma 4 E2B inference by 2.85 times on an A100 GPU using vLLM, with identical output tokens and no drop in quality. There's also a Gemma 4 Good Challenge on Kaggle for building products with positive impact.

Trying Gemma 4 in five steps

1. Decide where it has to run: a phone or device (E2B, E4B) or a computer (12B, 26B, 31B).
2. Try the model online first, in Google AI Studio or Google AI Edge Gallery.
3. Download the weights from Hugging Face, Kaggle, Ollama or LM Studio.
4. Check the Apache 2.0 licence with whoever signs off on software at your company.
5. Fine-tune only if the stock model misses your task.

Is Gemma 4 right for your business?

I'd look at Gemma 4 if you need a model that runs on your own hardware, works offline or keeps data off someone else's servers. I wouldn't start by downloading anything. Try it online first, check the licence with whoever signs off, and only fine-tune once you've seen the stock model miss a real task. If you just want a capable assistant and don't mind using Google's services, Gemini is the simpler route.

If your team is weighing up running an open model like Gemma 4 on its own hardware and wants the use cases, privacy and training questions settled first, that's the kind of thing I help with. I run AI training and workshops for teams, and help businesses with AI enablement, AI automation and GEO, which is about getting your business seen and cited in AI answers. Most of that work is with SaaS marketing teams, telecom companies and small businesses. Here's how we can work together.

Frequently asked questions

Is Gemma 4 free to use commercially?

Yes. Gemma 4 is released under the Apache 2.0 licence, which Google describes as commercially permissive. Check the licence with whoever signs off on software at your company.

What sizes does Gemma 4 come in?

As of November 2026 the DeepMind page lists E2B and E4B for phones and edge devices, and 12B, 26B and 31B for personal computers. The launch post named four sizes, without the 12B.

What's the difference between Gemma and Gemini?

Gemma is open weights that you download and run on your own hardware. Gemini is Google's model that you use through Google's services.

Where can I download Gemma 4?

The weights are on Hugging Face, Kaggle, Ollama, LM Studio and Docker. You can try the 31B and 26B in Google AI Studio and E4B and E2B in Google AI Edge Gallery first.

Can Gemma 4 run on a phone?

Yes. The E2B and E4B models run offline on phones, Raspberry Pi and NVIDIA Jetson Orin Nano devices, and take audio input as well as text and images.