Gemma 3n (preview)

2025-05-20 · Publisher : Google · Domain : Artificial Intelligence

Google unveils Gemma 3n in preview, an open model designed to run on phones, tablets and laptops, which fits in 2 to 3 GB of memory despite 5 to 8 billion raw parameters.

What changes

  • Mobile architecture shared with the next generation of Gemini Nano.
  • Per-Layer Embeddings (PLE): memory footprint equivalent to a 2 or 4 billion parameter model for raw sizes of 5 and 8 billion.
  • MatFormer: a 4 billion active-parameter model contains a 2 billion sub-model, with on-demand combination of intermediate sub-models.
  • Responses about 1.5 times faster on mobile than Gemma 3 4B, with better quality, according to Google.

### Specifications - Dynamic memory footprint: 2 GB and 3 GB. - Preview access: exploration in Google AI Studio (text) and local development with Google AI Edge (text and image); the changelog also cites `gemma-3n-e4b-it` in the Gemini API. - The full version arrives on June 26, 2025 (separate entry). ### Pricing Open model, no usage pricing of its own. ### Benchmarks A Chatbot Arena Elo chart appears in the post, with no figure in the text. ### Key takeaways A compact base for embedded and offline assistants, integrated into the Android ecosystem.

Links

Tags : Open Source, LLM, Google, Gemma, Embarqué