Gemini is not open source, but Google has released some parts of it

Google's Gemini is a large language model — the AI system behind Google's chatbot and other products — and it is not open source. Google does not publish the code that powers Gemini itself, and you cannot download it, modify it, or run it on your own computer. The weights and architecture remain proprietary, meaning only Google controls how the model works and what it can do.

However, the picture is more complicated than a simple yes or no. Google has released Gemma, a smaller family of models based on Gemini's technology, under an open source license. Gemma comes in different sizes (2 billion, 7 billion, and 27 billion parameters) and you can download it, study the code, modify it, and run it locally. This is a real open source offering, but it is not Gemini itself — it is a lighter-weight alternative that trades some capability for the freedom to use it however you want.

Key Takeaways

  • Gemini, Google's main AI model, is proprietary software you cannot download or modify, and you access it only through Google's services.
  • Gemma, a smaller open source model based on Gemini technology, is available for download and local use under the Gemma License.
  • Open source AI models let you run the software on your own hardware and modify it, but they often require technical knowledge and computing power.
  • If you need an AI model you can control completely, open source alternatives like Llama (from Meta) or Mistral are worth comparing to Gemma.

What open source actually means for an AI model

When people say an AI model is open source, they mean you get access to the trained weights — the numerical values the model learned during training — plus the code that runs it. This is different from open source software like Firefox or Linux, where you get the human-readable source code. With AI, the "source" is mostly the weights, which are large files (Gemma 7B is about 15 gigabytes) that you download and run through inference software.

Open source AI gives you real control: you can run it on your own server, integrate it into your own application, fine-tune it on your own data, and see exactly what it does. You are not sending your data to Google's servers or any company's servers. The trade-off is that you need the hardware to run it — a decent GPU or multiple CPUs — and the technical knowledge to set it up. Gemma's smaller sizes can run on a laptop with enough RAM, but the larger versions need serious hardware.

Why Google released Gemma but not Gemini

Gemini is Google's flagship product. It powers Google's chatbot, is integrated into Android phones, and is available through Google Cloud APIs that businesses pay for. Releasing Gemini as open source would undercut those revenue streams and give competitors direct access to Google's most advanced model. Companies rarely open source their most profitable technology.

Gemma serves a different purpose: it lets Google participate in the open source AI ecosystem, build goodwill with researchers and developers, and gather feedback on smaller models. It also lets Google say it supports open source AI without giving away its most valuable asset. This is a common pattern — companies release a capable but not cutting-edge version as open source while keeping the best version proprietary.

How Gemma compares to other open source models

Gemma is one option among several open source AI models. Meta's Llama 2 and Llama 3 are larger and often perform better on complex tasks, though they also require more computing power. Mistral's models are smaller and faster. Hugging Face hosts hundreds of open source models trained by researchers and companies, ranging from tiny (under 1 billion parameters) to very large.

The right choice depends on what you want to do. If you want to run something on a laptop or small server, Gemma 2B or 7B might be enough. If you need better reasoning or writing quality and have the hardware, Llama 3 70B is stronger. If you want to experiment without much setup, you can test these models on Hugging Face's website before downloading anything. None of them match Gemini's capabilities, but they are all genuinely open source — you own the code and can do what you want with it.

The license that comes with Gemma

Gemma is released under the Gemma License, which is Google's own license. It is not the GPL or MIT license you might see on other open source projects. The Gemma License lets you use the model for free, modify it, and distribute it, but it has restrictions: you cannot use it to develop competing AI models, you cannot use it for illegal purposes, and you have to include Google's attribution. These restrictions make it less "free" than licenses like MIT, but more open than proprietary software.

If you want a model with no restrictions at all, Llama 2 uses the standard Llama Community License, which is simpler. The point is that "open source" does not mean "no rules" — different open source projects have different rules, and you should read the license before you build something on top of it.

What you can and cannot do with Gemma

You can download Gemma, run it on your own computer or server, use it in your own application, and fine-tune it on your own data. You can sell a product that uses Gemma, as long as you follow the license terms. You cannot use Gemma to train a competing large language model — that is explicitly forbidden. You cannot use it for illegal purposes or to harm people. You must credit Google if you distribute Gemma or a modified version.

In practice, this means Gemma is suitable for building chatbots, search tools, content generation, code assistance, and other applications. It is not suitable if your goal is to create a new foundation model that competes with Gemini or other large models. For most people and small companies, these restrictions do not matter — you are free to build what you want.

How to access Gemma if you want to use it

You can download Gemma from Hugging Face, Google's official repository. Hugging Face hosts the model weights and provides tools to run them. You will need a machine with enough RAM (at least 8 gigabytes for the 7B model, more for larger versions) and some familiarity with Python or command-line tools. If you want to try it without downloading, Hugging Face lets you chat with Gemma in your browser, though this is slower and limited.

Google also offers Gemma through Google Cloud's Vertex AI service, where you can use it without downloading anything — you pay per request, similar to how you would use Gemini. This is easier if you do not want to manage your own hardware, but it is not open source in the sense of owning the code. The free download from Hugging Face is the truly open source route.

Frequently Asked Questions

Can I use Gemma to build a commercial product?

Yes. The Gemma License allows commercial use as long as you follow the terms — mainly that you cannot use it to train a competing large language model and must credit Google. Most commercial applications like chatbots, content tools, and code assistants are fine.

Is Gemma as good as Gemini?

No. Gemini is Google's most advanced model and performs better on complex reasoning, writing, and coding tasks. Gemma is smaller and faster but less capable. If you need the best performance and do not mind using Google's service, Gemini is stronger. If you need to run something locally or own the code, Gemma is the trade-off.

Do I need to be a programmer to use Gemma?

You need some technical knowledge — at least comfort with Python and command-line tools. If you want to use it without programming, you can access Gemma through Google Cloud's API or test it on Hugging Face's website. Running it locally requires more setup.

What is the difference between Gemma and Llama?

Both are open source models, but Llama (from Meta) is generally larger and more capable, especially Llama 3. Gemma is smaller and easier to run on modest hardware. Llama's license is simpler. The choice depends on your hardware and what you are building.

If Gemma is open source, can Google change it or take it away?

Google can stop releasing new versions of Gemma, but the versions already released remain open source and available. Once code is open source, it stays open source — Google cannot retroactively make it proprietary. However, future versions might have different terms.