Gemma, DeepMind's Family of Open Models — Omar Sanseviero, Google DeepMind

By ai.engineer

Categories: AI, Tools

Summary

Google DeepMind's Gemma 4 family achieves unprecedented capability-to-size efficiency, with models ranging 2-32B parameters that run on consumer GPUs, phones, and even Nintendo Switches. The new Apache 2 license and novel per-layer embedding architecture (E2B) enable practical on-device AI without API dependencies.

Key Takeaways

  1. Gemma 4's smallest models (2B parameters) run natively on Android, iOS, and Raspberry Pi with multimodal reasoning and agentic capabilities—eliminating infrastructure dependencies for edge deployment.
  2. The 27B mixture-of-experts model delivers high-speed inference at extremely low latency, enabling parallel processing (10 instances generating SVGs simultaneously at 100 tokens/second on consumer hardware).
  3. Apache 2 license shift from previous restrictive terms grants developers full flexibility and control—critical for enterprise adoption and open-source ecosystem participation.
  4. Per-layer embeddings (E2B architecture) replaces expensive computational layers with efficient lookup tables, enabling 4B parameter models to achieve 2B effective parameter efficiency gains.
  5. LM Arena benchmarks show Gemma 4 occupies the 'top-left corner'—small model size with high capability scores—demonstrating 2-year progression without proportional parameter scaling.

Topics

Transcript Excerpt

[music] >> All right. Hi everyone. It's cool here. So I'm super excited to give this talk because just 7 days ago we released Gemma 4. So before this conference, who here has heard about Gemma already? Okay, so most of you. Great. So Gemma is Google DeepMind's family of open models. Open models means that these are models that you can take, you can download, you can run in your own infrastructure, your own devices. You can fine-tune for your own use cases. So about a year ago we released Gemma 3...