Google Unveils Gemma 3 Open Models for Broad On-Device AI Deployment

March 12, 2025

Details

  • Google has introduced Gemma 3, a new family of lightweight open models, available in parameter sizes ranging from 1B to 27B, designed for widespread use on devices such as smartphones, laptops, and workstations.
  • Gemma 3 is developed in collaboration with hardware partners, optimized for NVIDIA GPUs, AMD ROCm environments, and Google Cloud TPUs.
  • The models support over 140 languages, offer a 128k-token context window, enable function calling, and include quantized versions to maximize efficiency on resource-constrained hardware.
  • The release also features ShieldGemma 2, a 4B-parameter image safety model, supporting safe deployment of AI systems and building on the ecosystem's momentum with 100 million downloads and over 60,000 community-created variants.
  • Google is launching an academic grant program and providing integration with platforms like Hugging Face and PyTorch to drive further accessibility and research.

Impact

The debut of Gemma 3 squarely positions Google against Meta’s Llama models, intensifying competition in the open AI landscape. Gemma 3’s enhanced efficiency and extensive language support may accelerate on-device AI adoption and development globally. With initiatives like ShieldGemma 2 and academic support, Google reinforces its focus on safe, responsible AI and broad platform integration.

Read more