Google Launches Gemma 4 12B for Local AI Workflows on Laptops

Google DeepMind has released Gemma 4 12B, a lightweight open model designed to run powerful AI workloads directly on consumer laptops. The release focuses on enabling local, agentic workflows through the Google AI Edge stack, eliminating the need for cloud services.

On-Device Capabilities#

The model-runtime combination unlocks several practical features:

• Autonomous data processing and analysis with rich visual outputs
• Code generation and execution for scripting tasks
• Full webpage building and tool use workflows
• Voice-driven text editing with advanced reasoning

New Applications#

Three applications now showcase Gemma 4 12B's capabilities:

Google AI Edge Gallery (macOS) A local AI showcase app demonstrating the model's coding strengths. Users can describe analytical goals in natural language—for example, requesting "use a python program to render a chart png to compare the top 10 girl names born in 2024 vs 2025"—and the model generates executable Python code that produces visualizations on-device.

Google AI Edge Eloquent (macOS) An on-device dictation and editing tool now featuring Voice Edit, a new capability that lets users transform text via voice commands. Users can highlight text and say commands like "restructure these notes into an executive summary" or "translate this into Hindi." According to Google, Gemma 4 12B delivers a "60%+ jump in overall quality" versus prior models, with improved instruction following and stricter scope adherence.

LiteRT-LM CLI The lightweight language model tool now includes a serve command, enabling users to run a local OpenAI-compatible LLM server directly from the terminal. This allows standard tools—such as Continue and Aider—to connect to a fully local endpoint powered by Gemma 4 12B.

Developer-Focused Design#

The release emphasizes privacy and efficiency. All data remains on-device, ensuring responsiveness without cloud dependencies. The approach aims to make agentic AI broadly accessible to everyday machines by pairing the model with optimized Google AI Edge infrastructure.