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·Oct 25·5 min read
Introducing Gemma 3n: The developer guide
TL;DR
Developers faced challenges in deploying powerful AI models on mobile devices due to performance limitations. Gemma 3n, built on a mobile-first architecture called MatFormer, enables advanced multimodal capabilities for edge devices.
✦ Why It Matters
Engineers can now deploy advanced AI models on mobile devices, enhancing application capabilities and user experiences.
Key Takeaways
- Experiment with Gemma 3n using Google AI Studio.
- Download model weights from Hugging Face or Kaggle.
- Utilize Mix-n-Match for custom model sizes.
- Participate in the Gemma 3n Impact Challenge for potential prizes.
How It Works
Gemma 3n's core innovation is the MatFormer architecture, which allows for nested transformer models that can dynamically adjust their size and performance. This enables developers to optimize models for specific hardware constraints while maintaining high performance.
Additionally, Per Layer Embeddings (PLE) improve memory efficiency by allowing only essential parameters to be loaded into the device's accelerator, enhancing overall model quality.