Google Drops Gemma-4-12B: One Model, Three Formats, Zero Encoders

Google has released Gemma-4-12B, a 12-billion-parameter open model that handles text, images, and audio in a single decoder-only system. The unified design ditches separate encoders, so all data goes straight into the transformer for faster local deployment. It’s part of the new Gemma 4 family from Google DeepMind, targeting consumer GPUs and workstations.

Google who also released  Gemma-4-E4B-it, built this model to give developers a multimodal tool that runs efficiently on a single high-end consumer GPU. By making audio and vision work natively, it avoids the overhead of separate processing pipelines. The open weights let privacy-conscious users keep all data on their own hardware.

Encoder-free design for local performance

Prosumers running local AI rigs can now handle audio tasks without extra models. Small agencies can process sensitive client data entirely offline on a single GPU. Hobbyists get a capable multimodal model that doesn’t require a server cluster.

What to know before you start

The model is integrated into Transformers and llama.cpp, with careful management of chat templates for thinking modes. For best results, place images before text and audio after text in your prompts. Audio is capped at 30 seconds and video at 60 seconds, with configurable image token budgets from 70 to 1120 to balance detail and speed.

“This unified approach to multimodality makes the model encoder-free, offering a deployment size that is perfect for consumer devices and streamlined local execution.” — Source: Hugging Face