Introducing Gemma 4 12B: a unified, encoder-free multimodal model
| Source: Google DeepMind Blog
Tags: Gemma 4, Google DeepMind, open-source LLM, multimodal, encoder-free, Apache 2.0, on-device AI
Google DeepMind releases Gemma 4 12B under Apache 2.0 — an encoder-free multimodal model with native audio inputs that runs on consumer laptops with 16GB of RAM, delivering benchmark performance close to the larger 26B MoE model.
Details
Google DeepMind's Gemma 4 12B is designed to bring multimodal AI to local hardware, targeting developers who want capable on-device agentic workflows without cloud dependencies. The model uses a novel encoder-free architecture: unlike traditional multimodal models that route vision and audio through separate encoders, Gemma 4 12B processes these inputs directly in the LLM backbone using a lightweight embedding module, reducing latency and memory overhead. The 12B model bridges the gap between Gemma 4's smaller E4B and the 26B Mixture of Experts model — achieving near-26B benchmark performance at less than half the memory footprint. It is the first mid-sized Gemma model to natively support audio inputs. Combined with Multi-Token Prediction (MTP) drafters, the model is positioned for both accuracy and speed in agentic use cases. Accessibility is a design priority: the model runs on 16GB of VRAM or unified memory, making it compatible with consumer-grade M-series MacBooks and mid-tier workstations. The Apache 2.0 license allows unrestricted commercial use. The Gemma family has now crossed 150 million cumulative downloads, indicating strong developer adoption. Community use cases highlighted include wearable robotic arms and enterprise AI security. The Gemma Cookbook provides integration examples for developers.