Google Unveils Gemma 3: The World’s Most Powerful Single-Accelerator Model

Google Announces Gemma 3 as 'World’s Best Single-Accelerator Model'

• Release and Context: On March 12, 2025, Google announced the release of Gemma 3, the latest addition to its family of open models for developers. This follows versions 1 and 2 released in February 2024 and May 2024, respectively.

• Gemma Family Overview: The Gemma family has seen over 100 million downloads in the past year, with 60,000 Gemma variants in what’s known as the 'Gemmaverse.' These models are designed to run efficiently on a variety of devices, from phones and laptops to workstations.

• Gemma 3 Features:

  • Performance: Built using the same research and technology as Gemini 2.0, Gemma 3 is available in four sizes: 1B, 4B, 12B, and 27B parameters.
  • Superior Single-Accelerator Performance: Google claims it outperforms models like Llama-405B, DeepSeek-V3, and o3-mini in the LMArena benchmark.
  • Advanced Capabilities: The 4B+ sizes offer enhanced text and visual reasoning capabilities, supporting image, text, and short video analysis. It has a 128k-token context window and supports over 35 languages out-of-the-box, with pre-trained support for over 140 languages.
  • AI-Driven Workflows: Supports function calling and structured output to automate tasks and build agentic experiences.
  • Quantized Models: Officially quantized versions reduce model size and computational requirements while maintaining high accuracy.

• Safety Features:

  • Image Safety Checker: Includes a 'ShieldGemma 2' for detecting dangerous, sexually explicit, and violent content in images.
  • Data Governance: Ensures alignment with safety policies through fine-tuning and robust benchmark evaluations.

• Availability: Gemma 3 can be tried in Google AI Studio, with model downloads available on Kaggle or Hugging Face.