Google’s new Gemma 4 models bring sophisticated computing capabilities to low-power devices

Google LLC expands the power of open-weighted artificial intelligence models with the release of Gemma 4, its most advanced “open” model family to date.

Built on the same architectural foundation as the Gemini 3, the models are designed to handle complex computing tasks and support autonomous AI agents operating locally on low-power devices such as workstations and smartphones.

With Gemma 4, Google DeepMind researchers Clement Farabet and Olivier Lacombe say, they were able to extract “a lot of intelligence per parameter,” allowing them to punch well above their weight class. For example, the 31B Dense variant is currently ranked third among open models on the industry-standard Arena AI Text leaderboard.

The Gemma 4 models come in four flavors: the 2B Active, the 4B Efficient, the 26B Professional Blend model and the 31B Dense model. The smaller “Active” models are designed for lightweight hardware use cases such as Android smartphones or Raspberry Pi computers, the researchers said. Meanwhile, the 26B MoE model has a clever trick because it only uses 3.8 billion parameters for inference operations, allowing it to perform at high speed without sacrificing the deep knowledge base of larger models.

Farabet and Lacombe explain that each of the Gemma 4 models is better suited for using AI agents. While previous iterations of Gemma forced developers to adapt their architecture to interoperability with other software tools, Gemma 4 models have native support for function calling and structured output for JavaScript Object Notation. This means developers can use them to power autonomous agents that interact with third-party tools and use them in multi-step applications.

All four models have the ability to process images and videos, the E2B and E4B variants go further with native audio input support, allowing real-time speech understanding directly on the device. Google also increased the context window for the models, up to 128K for the smallest models and 256K for the two largest. This means that developers will be able to upload an entire codebase or large sets of documents with a single command.

Each model is made available under a valid Apache 2.0 license, which removes many of the commercial restrictions imposed on other AI models, making them a good choice for developers building business applications, Google said. They can be accessed directly through Google Cloud, and are available with their open weights on Hugging Face, Kaggle and Ollama.

The release underscores Google’s ambitions to dominate the “AI space” industry. Because even the largest Gemma 4 models are small enough to run on a single graphics processing unit, that makes them ideal for edge use cases and applications where low latency and digital sovereignty are critical, said Holger Mueller, an analyst at Constellation Research.

“Google is building its leadership in AI, not only by pressing Gemini, but also by opening models with the Gemma 4 family,” he said. “This is important for building an ecosystem of AI developers, and will help the company tap into more functional and vertical use cases in different device form factors. Google set the bar high with its previous release of Gemma 3, so there’s a lot to look forward to with this release.”

Image: Google