Google has introduced Gemma 4, a new open-source model: Here’s how to test it

On Thursday, Google unveiled the newest iteration of its open AI model, Gemma 4. Importantly, Gemma 4 is a completely open-source model licensed under Apache 2.0, which is generally uncommon for cutting-edge models.

Open models are capable of being executed locally on users’ devices, and Google claims that Gemma 4 can operate on “billions of Android devices” as well as certain laptop GPUs.

This open-source license establishes a basis for total developer flexibility and digital sovereignty; it gives you full control over your data, infrastructure, and models,” a blog post from Google states. “It enables you to develop freely and deploy securely in any setting, whether on-premises or in the cloud.

Many individuals are probably familiar with Google’s well-known Gemini AI model, largely due to the widespread AI chatbot that has been incorporated into numerous Google products.

Gemma is a large language model (LLM) developed using the same technology and research that Google DeepMind employed to create Gemini 3.

Google refers to Gemma 4 as its “most capable” open AI model to date.

Gemma vs. Gemini?

So, what distinguishes Gemma from Gemini?

Gemini is Google’s exclusive subscription-based AI product and the designation for Google’s suite of multimodal AI models. It has been incorporated into nearly all of Google’s essential offerings, such as Google Search, Gmail, Google Docs, and Google Cloud.

On the other hand, Gemma 4 is an open AI model, which means that both the code and the data it utilizes for training are accessible to its users. Users can operate Gemma AI models on their local hardware, even in the absence of an internet connection. Anyone is able to download Gemma 4 and execute it on their device at no cost. These open AI models ensure a more private and secure experience, as none of the conversations, uploaded documents, or responses are disclosed to any third party.

Developers can utilize open AI models like Gemma 4 to incorporate AI into their applications without incurring any ongoing subscription fees.

What is Gemma 4?

Gemma 4 introduces advanced features to Google’s open AI model lineup.

As per Google’s announcement, Gemma 4 is now equipped with advanced reasoning capabilities, which encompass multi-step planning and complex logic. Google claims to have achieved “significant enhancements in mathematical and instruction-following benchmarks that necessitate it” with Gemma 4.

Additionally, Gemma 4 supports processes essential for agentic workflows and provides localized AI coding assistance. Furthermore, Gemma 4 is capable of processing audio and video for speech recognition and interpreting visual data such as charts.

Gemma 4 is offered in four sizes, determined by the number of weights utilized to power the model: two billion, four billion, 26 billion, and 31 billion.

Hugging Face indicates that these open-weight models are available in both pre-trained and instruction-tuned versions, providing even greater flexibility for developers.

According to Google, the AI model has been trained on over 140 languages and features a context window of up to 256,000 tokens. (However, the smaller E2B and E4B variants have a context window of 128,000.)

Gemma 4 is now open and open source

Now, the term open does not equate to open source in the realm of AI models.

Earlier versions of Gemma were open-weight (indicating that the training datasets are publicly accessible) but were still subject to Google’s terms, even though users could download the model onto their devices. While users had the ability to modify the local LLM, they were still required to adhere to Google’s regulations regarding its use and redistribution.

With Gemma 4, Google has officially made the model both open and open source.

Google is distributing Gemma 4 under the widely recognized open source software license Apache 2.0.