Google launches Gemma 4, multimodal and local AI

Google has introduced Gemma 4, a new family of open artificial intelligence models that can work with text, images, video and, in its smaller versions, audio. The models are released under the Apache 2 license and designed to run everywhere from servers to personal computers, browsers and resource-constrained devices.

The family includes four sizes, with base versions and versions tuned to follow instructions:

• Gemma 4 E2B: 2.3 billion effective parameters and a 128,000-token context window.
• Gemma 4 E4B: 4.5 billion effective parameters and a 128,000-token context window.
• Gemma 4 31B: 31 billion parameters and a 256,000-token context window.
• Gemma 4 26B A4B: mixture-of-experts model that activates 4 billion parameters out of a total of 26 billion, with a 256,000-token context window.

A token is a unit of text that the model processes, such as a word or part of one. A context window of 128,000 or 256,000 tokens allows it to analyze long documents, extended conversations or large amounts of code without having to split them up as much.

A small model with broad capabilities

Gemma 4 can describe images, extract text from documents, detect objects and locate specific elements on a screen. In published tests, the model identifies a button labeled "view recipe" and returns its coordinates in JSON format, which is useful for agents that can interact with applications.

It can also generate code from a screenshot. For example, it receives an image of a webpage and produces the HTML needed to recreate it.

The smaller E2B and E4B models accept text, images, video and audio. The larger models process images and video, although in video tests they do not include the audio track. The family also supports speech transcription and understanding.

There is an important limitation: audio training focused on speech, not music or ambient sounds. In a test involving a concert, E2B invented part of the audio performance, while E4B correctly described the musical speech and the visual scene. You should not treat its responses as a reliable transcription of every kind of sound.

Performance with far fewer parameters

According to estimates shared by Hugging Face, the dense 31B model reaches a score of 1452 on LMArena, a ranking based on comparisons between model responses. The Gemma 4 26B A4B model scores 1441 while activating only 4 billion parameters for each response.

The comparison matters because, according to the publication, those scores are close to models such as GLM-5 or Kimi K2.5, but with approximately 30 times fewer parameters. These are estimated figures based on text tests, not a universal measurement of all their capabilities.

The architecture combines local attention, which focuses on nearby segments, with global attention to connect distant information. It also reuses part of its internal memory during generation. The goal is to reduce memory and computing requirements, especially when the model works with long contexts or runs on a local device.

Available in more tools from day one

Gemma 4 is integrated into several open-source technologies, including:

• Hugging Face Transformers and TRL.
• llama.cpp, with support for applications such as LM Studio, Jan and coding agents.
• MLX for Apple Silicon devices.
• transformers.js to run it in the browser through WebGPU.
• mistral.rs to use it from Rust.
• ONNX for different types of hardware.

This makes it possible to use a model such as Gemma 4 E2B on a personal computer or in a browser, while the larger versions can be deployed on servers with more memory. The Apache 2 license also makes it easier to integrate, modify and fine-tune for your own projects, although each use must comply with the license's conditions and applicable policies.

For you, the main change is that an AI capable of reading images, understanding speech and using tools can work without necessarily sending all your data to an external service. That opens up possibilities for private assistants, local automation, document analysis and agents that operate on computers or connected devices.

Gemma 4 does not eliminate the errors of multimodal models. It can confuse locations, misinterpret audio or invent details. What to watch now is whether its availability across so many devices turns these capabilities into real, reliable applications, rather than just technical demonstrations.