Google DeepMind launches Gemma 4: AI models for advanced reasoning and AI agents

Google DeepMind has introduced Gemma 4, a new generation of open AI models designed for advanced reasoning and autonomous AI workflows. The release represents Google’s latest push in the open-model ecosystem, offering developers more powerful AI tools while allowing broad use under the Apache 2.0 open-source licence.

According to Google DeepMind’s official blog, Gemma 4 is designed to deliver higher intelligence with fewer parameters, making it more efficient for real-world applications.

Four model sizes designed for different workloads

Google released Gemma 4 in four model configurations, each designed for specific computing needs.

The models include:

E2B (effective 2 billion parameters) - lightweight model for edge devices

E4B (effective 4 billion parameters) - improved performance for mobile environments

26B Mixture-of-Experts (MoE) - optimised for complex reasoning tasks

31B Dense model - the most powerful version in the Gemma 4 series

According to benchmark rankings, the 31B model currently holds the third position among open AI models on the Arena AI leaderboard, while the 26B model ranks sixth, outperforming models estimated to be 20 times larger.

A Google DeepMind research leader wrote in the official announcement that “Gemma 4 delivers an unprecedented level of intelligence-per-parameter.”

Built for agentic AI systems

One of the key goals of Gemma 4 is to support agentic AI systems, which can perform tasks autonomously rather than simply generating text responses.

The models include several capabilities designed for AI agents, such as:

Native function calling
Structured JSON output
Support for system commands
Integration with external tools and APIs

These features allow developers to build AI systems that can analyse data, trigger workflows and interact with digital tools.

Multimodal AI with support for 140 languages

Gemma 4 is also designed as a multimodal AI system, meaning it can process different types of data.

The models support:

Text
Images
Video

Additionally, the smaller E2B and E4B versions include native audio input for speech recognition.

Google said the models support more than 140 languages, allowing developers to build AI applications for global users.

Another notable upgrade is the 256,000-token context window for larger models. This enables developers to process long documents, full research papers or even entire code repositories in a single prompt.

AI models designed for mobile and edge devices

Unlike many large AI systems that require powerful servers, some Gemma 4 models are designed to run directly on devices. The E2B and E4B models are optimised for mobile phones, Internet-of-Things devices and edge computing systems.

Google said these models can run offline on devices such as smartphones, Raspberry Pi computers and NVIDIA Jetson Orin Nano, offering near-zero latency.

The company also collaborated with Qualcomm and MediaTek to improve mobile performance.

A growing ecosystem around Gemma models

Since the first Gemma models were introduced, adoption has grown rapidly. Google reports that Gemma models have been downloaded more than 400 million times, leading to the creation of over 100,000 community-built variants, often referred to as the “Gemmaverse”.

Developers can access Gemma 4 through platforms such as:

Hugging Face

Kaggle

Ollama

Google AI Studio

What this means for developers and the AI industry

Gemma 4 reflects a broader shift in the AI industry toward open models that developers can customise and deploy across different environments.

By offering multiple model sizes and device compatibility, Google DeepMind aims to support both large-scale enterprise systems and smaller applications running on personal devices. As AI development moves toward autonomous systems capable of executing tasks, models like Gemma 4 may become an important foundation for building next-generation AI agents.