We’re launching Lyria 3.5 in Google Flow Music, with advances across musicality, lyrics, vocals, and creative control
deepmind.google·6d ago
AI Engineering
DeepMind Blog·Jun 9·5 min read
Introducing Gemma 4 12B: a unified, encoder-free multimodal model
TL;DR
Multimodal models, which process different types of data like text and images, often rely on complex architectures. Gemma 4 12B is a new unified, encoder-free multimodal model designed to run state-of-the-art agents locally.
✦ Why It Matters
Engineers can leverage Gemma 4 12B to create more efficient multimodal applications without the complexity of traditional architectures.
Key Takeaways
- Run Gemma 4 12B locally on laptops with 16GB of RAM.
- Utilize the unified architecture for efficient audio and visual input processing.
- Access the model under an open-source Apache 2.0 license.
- Explore the developer documentation for integration with existing tools.
How It Works
Gemma 4 12B employs an encoder-free architecture, allowing audio and visual inputs to be processed directly by the language model. For vision, it uses a lightweight embedding module instead of a traditional encoder, while audio is projected into the same dimensional space as text tokens, streamlining the processing pipeline.