Gemma 4 has been released — and now AI agents can be run directly on a laptop

Key Features of Gemma 4

The models are designed to run on user-owned hardware—phones, laptops, and desktop PCs. For the first time, release under the Apache 2.0 license allows free use, modification, and distribution.

Gemma 4 is oriented toward the era of agents:

— handles complex logic

— multi-step planning

— workflows.

The large model supports a context window of up to 250,000 tokens for analyzing codebases and multi-turn scenarios. Built-in tool support enables agents to plan and act autonomously.

Gemma 4 Model Lineup

The 26B Mixture of Experts and 31B Dense models deliver cutting-edge intelligence on PCs. Run local reasoning and coding pipelines without sending data to the cloud.

— 26B MoE with 3.8 billion active parameters runs extremely fast

— 31B Dense is optimized for text generation quality.

The 2B and 4B models are built for mobile devices and IoT with maximum memory efficiency. Support for real-time audio and video, native operation in over 140 languages.

Testing the 2B Model

Oliver tested the efficient 2B model on a multilingual and agentic task. This was a real-world test of Gemma 4’s capabilities in everyday scenarios—a combination of understanding a French query, searching for information, and responding in English. “Gemma, can you find a French restaurant in San Francisco? Answer me in English.”

The model handled the task, executing a multi-step request without delays. It didn’t just translate; it applied agentic logic: found relevant data and provided an accurate response.

Security and Integration

Gemma 4 undergoes the same rigorous security protocols as Google DeepMind’s proprietary models.

This provides enterprises and developers with a reliable foundation for building applications and solutions. Gemma 4 is fully compatible with tools you already know. Model weights are available for download right now—start experimenting today.