Google launches Gemma 4, its open AI models

Google has introduced Gemma 4, a new family of open artificial intelligence models designed to reason, use tools and run directly on computers, phones and other devices. It comes in four sizes and under an Apache 2.0 license, which allows developers to adapt and deploy the models with few restrictions.

The company says Gemma has already reached more than 400 million downloads and that the community has created more than 100,000 variants. Gemma 4 aims to bring advanced capabilities to devices that cannot always run massive cloud-based models.

Four sizes for different devices

The family includes two small models, E2B and E4B, and two higher-capacity models: a 26B model based on a Mixture of Experts architecture and a dense 31B model. In Mixture of Experts models, only part of the parameters are activated for each response, reducing cost and latency.

According to Google, the 31B currently ranks third among open models in Arena AI’s text leaderboard, while the 26B ranks sixth. The company also says the 26B can outperform models up to 20 times larger in certain evaluations. These are benchmark results, not a guarantee of identical performance on every task.

The large models are designed for personal computers and workstations. Their unquantized weights in bfloat16 format fit on an 80 GB NVIDIA H100 GPU, while quantized versions can run on consumer GPUs. In practice, this makes it possible to build coding assistants or local agents without sending every request to an external server.

The E2B and E4B models prioritize use on phones and small devices. They can run offline on Android phones, Raspberry Pi and devices such as the NVIDIA Jetson Orin Nano, with lower memory and battery consumption. They also include audio input for recognizing and understanding speech.

What Gemma 4 can do

Google has expanded Gemma beyond basic chat. The family offers:

• Multi-step reasoning for logic and math problems, as well as following instructions.
• Tool use, through function calls and structured JSON responses. This lets you connect the model to APIs and run tasks according to a defined workflow.
• Code generation for working locally from a computer.
• Image and video understanding, including recognizing text in images and reading charts.
• Context windows of up to 128,000 tokens in the models for devices and up to 256,000 in the large models. This makes it possible to analyze lengthy documents or entire code repositories in a single request.
• Native support for more than 140 languages.

For you, the most visible change may be that some AI features can run directly on the device. A coding assistant could work offline, an app could analyze an image without uploading it to the cloud and an agent could respond with less delay. The trade-off is that the result will depend on the hardware and on how each developer adapts the model.

Available to developers

Gemma 4 can be tested in Google AI Studio and Google AI Edge Gallery. The weights are available through Hugging Face, Kaggle and Ollama, along with tools such as llama.cpp, vLLM, MLX, Transformers and LM Studio.

The Apache 2.0 license allows you to use, modify and deploy the models in commercial projects. It also gives you more control over data and infrastructure, although the responsibility for evaluating the model’s safety, accuracy and behavior falls on whoever implements it.

Gemma 4 reinforces Google’s strategy of offering two paths: Gemini for people who want a ready-to-use proprietary service and Gemma for those who need to download, fine-tune and run the model under their own control. The key question now is how much of that performance reaches real-world applications, especially on phones and computers with limited resources.