Google DeepMind has launched Gemma 4, its latest generation of open AI models, and the numbers are turning heads. Released under the Apache 2.0 open-source licence — meaning developers can use, modify, and build on it freely — Gemma 4 is designed to pack more intelligence into fewer parameters than its predecessors. For the AI development community, this is a significant release worth paying close attention to.

Four Models, One Family — Each Built for a Different Job

Gemma 4 isn’t a single model — it’s a suite of four, each calibrated for different computing environments and use cases. The smallest is the E2B (effective 2 billion parameters), a lightweight model built specifically for edge devices where resources are tight. Step up to the E4B (effective 4 billion parameters) and you get meaningfully better performance, optimised for mobile environments.

For heavier lifting, the 26B Mixture-of-Experts model is built for complex reasoning tasks, while the 31B Dense model sits at the top of the family as the most powerful version in the Gemma 4 lineup. According to benchmark rankings on the Arena AI leaderboard, the 31B currently holds third place among all open AI models globally — while the 26B sits at sixth, despite outperforming models estimated to be 20 times its size. That’s a remarkable efficiency ratio.

“Gemma 4 delivers an unprecedented level of intelligence-per-parameter.”

— Google DeepMind research leader, official announcement

These Models Are Built to Act, Not Just Answer

A lot of AI models are good at generating text. Gemma 4 is designed to go further — supporting agentic AI systems that can actually perform tasks autonomously. Think less chatbot, more digital assistant that can take action on your behalf.

To make that possible, Gemma 4 includes native function calling, structured JSON output, support for system commands, and integration with external tools and APIs. In practice, this means developers can build AI systems that don’t just respond to prompts — they can analyse data, trigger workflows, and interact with other software tools in real time. It’s the kind of infrastructure that next-generation AI agents are built on.

Multimodal, Multilingual, and Built for the World

Gemma 4 can process text, images, and video — and the two smaller models (E2B and E4B) also include native audio input for speech recognition. That multimodal capability, combined with support for more than 140 languages, means developers can build genuinely global applications without stitching together multiple separate tools.

The larger models also come with a 256,000-token context window — a substantial upgrade that allows them to process long documents, full research papers, or entire code repositories in a single prompt. For developers working with complex, data-heavy applications, that’s a game-changer.

Running AI on Your Phone — Without an Internet Connection

One of the more striking aspects of Gemma 4 is how far down the hardware stack it can reach. The E2B and E4B models are designed to run directly on smartphones, Raspberry Pi computers, IoT devices, and systems like the NVIDIA Jetson Orin Nano — offline, with near-zero latency. Google worked with both Qualcomm and MediaTek to optimise mobile performance, making on-device AI a practical reality rather than a distant promise.

For applications where privacy, connectivity, or speed are concerns, running a capable AI model entirely on-device is a meaningful step forward.

400 Million Downloads and a Growing Community

Gemma models have been downloaded more than 400 million times since the first generation launched, spawning over 100,000 community-built variants — a collective the developer community has started calling the “Gemmaverse.” That scale of adoption signals genuine trust and momentum in the open-model ecosystem.

Gemma 4 is available now through Hugging Face, Kaggle, Ollama, and Google AI Studio, making it accessible across the platforms developers already use.

What This Means for the AI Industry

Gemma 4 is part of a broader shift happening across the AI landscape — away from closed, proprietary systems and toward open models that developers can inspect, customise, and deploy on their own terms. By offering a range of model sizes that scale from a pocket-sized smartphone app all the way up to enterprise-grade reasoning systems, Google DeepMind is making a clear bet: the future of AI isn’t one giant model running in the cloud. It’s intelligent systems running everywhere, on everything.

Whether you’re building a mobile app, a research tool, or an autonomous AI agent, Gemma 4 is designed to be the foundation you build it on.