Introducing Gemma 4 12B: A Unified, Encoder-Free Multimodal Model

Date:June 3, 2026

Google DeepMind is proud to introduceGemma 4 12B, a new model designed to bring high-performance multimodal intelligence directly to consumer laptops. Bridging the gap between the edge-friendly E4B and the advanced 26B Mixture of Experts (MoE), Gemma 4 12B packages powerful capabilities into a significantly reduced memory footprint. Notably, it is the first mid-sized model in the Gemma series to feature native audio inputs.

Key Features and Capabilities

Gemma 4 12B is engineered to deliver agentic multimodal intelligence on everyday hardware without sacrificing speed or reasoning power. Its standout features include:

• Novel Unified Architecture:The model eliminates the need for separate multimodal encoders, allowing vision and audio inputs to flow directly into the Large Language Model (LLM) backbone.
• Advanced Reasoning:Benchmark performance approaches that of the larger 26B model, unlocking powerful multi-step reasoning and agentic workflows.
• Laptop Ready:Optimized for local execution on consumer laptops equipped with just 16GB of VRAM or unified memory.
• Open and Accessible:Released under an Apache 2.0 license, ensuring broad support across the developer ecosystem.
• Drafter-Ready:Equipped with Multi-Token Prediction (MTP) drafters to significantly reduce latency during inference.

A Uniquely Efficient, Unified Architecture

What distinguishes Gemma 4 12B is its streamlined approach to processing visual and audio data. Traditional multimodal models typically rely on separate encoders to translate images and audio before passing representations to the language model. These split encoders often add latency and increase memory usage. To address this, Gemma 4 12B utilizes anencoder-free architecturethat integrates audio and vision inputs directly.

Vision Processing

Gemma 4 12B replaces the traditional vision encoder with a lightweight embedding module. This module consists of:* A single matrix multiplication.* Positional embedding.* Normalizations.

This design allows the LLM backbone to take over visual processing directly, reducing overhead.

Audio Processing

The audio processing pipeline has been simplified further. The model removes the audio encoder entirely, instead projecting the raw audio signal directly into the same dimensional space as text tokens. This enables native audio input capabilities within the model.

Performance and Accessibility

Gemma 4 12B delivers state-of-the-art agent performance locally. It achieves benchmark results nearing the larger 26B MoE model while maintaining less than half the total memory footprint. This efficiency makes it possible to run powerful multimodal and agentic experiences directly on a user's machine.

Getting Started

Developers and researchers can access Gemma 4 12B through several channels:

• Try it Yourself:Experiment with the model using a few clicks in the provided interface.
• Download Weights:Access pre-trained and instruction-tuned checkpoints directly from the official repository.
• Integrate & Learn:Review comprehensive documentation to understand implementation details.
• Development Tools:Implement local inference pipelines using preferred development tools.
• Gemma Skills:Utilize official resources to support agents in building with the latest Gemma advancements.
• Deployment:Spin up production endpoints using Google Cloud or deploy via other preferred methods.

With Gemma 4 12B, Google DeepMind continues to push the boundaries of what is possible on local hardware, bringing advanced multimodal reasoning to the edge.