Google Debuts Gemma 4 for Edge Devices and AI Agents

News

1. Sarvam Seeks to Raise $350 Million at $1.55 Billion Valuation

2. The Five Skills LinkedIn Says Will Keep Humans Ahead of AI

3. OpenAI Moves Into Media With TBPN Deal

4. Google Debuts Gemma 4 for Edge Devices and AI Agents

5. India Defers SIM-Binding Rules for Messaging Apps

6. India Data Center Investment Seen Topping $54 Billion This Year

Google on Thursday introduced Gemma 4, the latest iteration of its open model family, signaling an effort to pair proprietary frontier systems with more accessible models that developers can run, fine-tune and deploy across a broad range of hardware.

Google said the Gemma family has been downloaded more than 400 million times and has spawned more than 100,000 variants since its debut.

In a post on X, CEO Sundar Pichai said, “Gemma 4 is here, and it’s packing an incredible amount of intelligence per parameter.”

Gemma 4 is being released in four sizes, 2B, 4B, 26B (Mixture of Experts), and 31B, designed to run across a wide range of hardware, from mobile devices to high-end GPUs. 

The company claims the models can handle more complex reasoning and multi-step workflows compared to earlier versions, while remaining lightweight enough for local deployment.

This emphasis on “intelligence-per-parameter” reflects a broader industry trend: building models that deliver competitive performance without the infrastructure costs associated with frontier-scale systems.

Google said its larger models rank among the top open models on benchmark leaderboards, though such rankings can vary depending on methodology and evaluation conditions.

A key shift in Gemma 4 is its push toward on-device AI. The smaller variants are optimised for mobile and edge hardware, including smartphones and IoT devices, with support for offline processing and low-latency tasks.

The models also include multimodal capabilities, processing text, images, video, and, in some cases, audio, along with longer context windows of up to 256K tokens in larger versions.

This positions Gemma 4 for use cases such as local code generation, document analysis, and autonomous “agentic” workflows that can interact with APIs and tools without constant cloud access.

Google is releasing Gemma 4 under an Apache 2.0 license, allowing developers to modify and deploy the models commercially. The move aligns with the company’s strategy of maintaining a foothold in the open ecosystem while continuing to develop its proprietary Gemini models.

The company said Gemma 4 shares underlying research with its Gemini systems, suggesting a dual-track approach: open models for flexibility and reach, alongside closed models for enterprise-scale deployments.

Early ecosystem partners, including Hugging Face, said they would support the models from launch. 

“The release of Gemma 4 under an Apache 2.0 license is a huge milestone. We are incredibly excited to support the Gemma 4 family on Hugging Face on day one,” said Clément Delangue, co-founder and CEO, Hugging Face.