Late last year, Google introduced Gemini 3 Pro, presenting it as a notable advance in its line of proprietary language models. The step forward was neither abrupt nor overstated, but part of a gradual and methodical progression. What distinguished it was not merely scale, but the refinement of ideas already in motion. Now, the company has decided to carry some of that same work beyond its own walls, making a portion of it available to the open-source community under the name Gemma 4.

Four Model Sizes Designed for Different Computing Needs

This new release is divided into four separate models, each defined by the number of parameters it contains. The smaller variants, with two billion and four billion “effective” parameters, are intended for use on devices where computing power is limited, such as smartphones. The larger offerings, by contrast, are built for systems with greater capacity: a twenty-six-billion parameter mixture-of-experts model, and a thirty-one-billion parameter dense model, each suited to more demanding environments.

Parameters, put plainly, are the adjustable elements within a model that influence how it produces an answer. A greater number of them tend to allow for more capable and nuanced responses. Yet this apparent advantage carries its own requirement. Models of larger size do not operate lightly; they depend on stronger hardware and consume more resources. The balance, therefore, is a familiar one. It’s between what can be achieved and what must be spent to achieve it.

Google Claims High Intelligence Per Parameter

With Gemma 4, Google states that it has achieved what it calls an unusual degree of intelligence for each parameter. The claim is direct and leaves little room for ornament. In support of it, the company refers to the performance of its larger models. The thirty-one-billion and twenty-six-billion versions secured third and sixth place on Arena AI’s text rankings. In doing so, they stood above systems far larger than themselves, some by a wide margin.

Multimodal Capabilities Expand Practical Use Cases

Each model can work with both images and video. This allows them to handle tasks such as reading text from pictures with a fair degree of ease. The smaller versions go further. They can take in audio and make sense of spoken language. This gives them a wider field of use, especially on devices where space and power are limited.

Google also notes that these models can produce code without needing a constant connection. This makes it possible to work in isolation, without relying on external systems. In addition, the models have been trained across more than one hundred and forty languages. The aim, it seems, is coverage rather than selectivity, and reach rather than excess.

Apache 2.0 License Brings Greater Developer Freedom

Google just made a change with its Gemma 4 models. They used to have a license that was firmer. Now they are using the Apache 2.0 license. This means that people can do a lot more with the models. They can change them. Use them in whatever way they want.

Google says this change will help developers work on their own. They will be able to control their data and systems and use the models without a lot of rules to follow. Developers can use the models on their computer or on the internet depending on what they need to do. Google says this will give developers freedom to work with the Gemma 4 models.

Availability Across Popular AI Platforms

If someone wants to try the Gemma 4 models it is not that hard. The files for the Gemma 4 models are available on platforms, like Hugging Face, Kaggle and Ollama. So the Gemma 4 models are ready to use, it just depends on what people want to make with them. So, the Gemma 4 models are available and people can start using them to build whatever they want.

Final Words

Google’s Gemma 4 does not feel like a change. It feels like a version of something you already like. It is the same idea but it is a bit better. There are different versions of Google’s Gemma 4. So whether you are using a laptop or working on a project in the cloud there is probably a Google’ s Gemma 4 that is right for you. The fact that Google’s Gemma 4 uses the Apache 2.0 license is a big deal. It is not a big deal in a bad way. It is like Google is saying you can do what you want with Google’s Gemma 4. You do not have to follow a lot of rules. This makes it more fun to try things with Google’s Gemma 4.

Google’s Gemma 4 can handle a lot of things. It can handle types of data and different languages. You can even use it to code without the internet. So it’s like a toolbox. It is for trying things not for showing off. Google’s Gemma 4 is not loud or flashy. It is just really good. It might not change everything away but it gives developers better tools and more freedom.