Is Gemma 4 free? Yes, the weights are Apache 2.0

Yes, Gemma 4 is free. Google released the model weights under the Apache 2.0 license, which makes Gemma 4 open source in the sense the Open Source Initiative uses, and lets anyone download it, run it and use it commercially at no charge. Google announced the models on April 2, 2026.

The "free" has two parts, and they are easy to mix up. The weights cost nothing. Running them still costs whatever hardware or cloud time you use, and Google's hosted options have their own terms.

Is Gemma 4 open source?

Mostly yes. Google's open source team says Gemma 4 is the first in the Gemma family released under the OSI-approved Apache 2.0 license. Earlier Gemma versions shipped under Google's own Gemma terms, so this is a real change. The Hugging Face model card for Gemma 4 31B lists the license as Apache 2.0 as well.

Apache 2.0 allows modification, reuse and commercial products. You can read the full text on the Apache Software Foundation site.

One caveat for strict readers. The usual open source definition covers code and its license. Google publishes the weights, and the model card describes them as open weights. The post does not say that the training data is open. Whether that counts as "open source" is a long-running argument, and the license itself is not in doubt.

Where you can download Gemma 4

According to Google's announcement, the weights are on Hugging Face, Kaggle and Ollama. Google's developer documentation points to Kaggle and Hugging Face, and says official quantization-aware training checkpoints are available there too. Those are versions trained to run at lower precision, so they use less memory.

The announcement lists four models:

A mixture-of-experts model runs only part of its network for each token, so it is faster than its size suggests. The developer documentation also lists a 12B model, which the announcement does not mention, so check the docs for the current lineup.

What it costs to run

Nothing is charged for the weights. What you pay for is compute. Google's documentation gives approximate memory needed just to load the weights in 16-bit precision:

The 26B mixture-of-experts model activates about 4B parameters per token but still has to load all 26B, which is why its memory need is much higher than a 4B model's. The docs also list a 12B model at 26.7 GB. Quantized checkpoints need less. Google does not give a dollar cost for running any of these on your own machine or in the cloud.

Gemma 4 in the Gemini API: free tier only

If you would rather call Gemma 4 through Google's hosted API, the Gemini API pricing page has a Gemma 4 section. As of October 30, 2026, it lists input, output and context caching as free of charge on the free tier. Every paid-tier line for Gemma 4 reads "Not available."

That means there is no paid Gemma 4 plan on the Gemini API, and the page gives no rate limits in the part we read. It also says that on the free tier, Google uses your content to improve its products, and on the paid tier it does not. Since Gemma 4 has no paid tier, anything you send through the free API falls under the first rule. If that matters for your data, download the weights and run the model yourself.

The announcement also says you can try the 31B and 26B models in Google AI Studio and deploy on Google Cloud through Vertex AI, Cloud Run or GKE. It does not say what those cost, and Google Cloud deployments will bill for the infrastructure you use.

What to check before you ship

Apache 2.0 is permissive, but it still comes with conditions such as keeping the license notice and stating changes. Read the license before building a product on it. Google's developer documentation also links a prohibited use policy, so confirm whether it applies to the version you download.

Prices and plans change. Check the pricing page and the model card on the day you decide.

For a wider look at what Google has released, see our models section.