Highlights:

  • Google claims Gemma 3 outperforms larger models like Llama-405B, DeepSeek-V3, and OpenAI’s o3-mini in LMArena’s preliminary evaluations, thanks to its efficient design.
  • Gemma is built on the same technical research as Google’s Gemini model, the company’s most advanced and powerful AI to date.

Google LLC has launched next generation Lightweight Gemma AI Models, designed to run on a single graphics processing unit, as part of its ongoing effort to make artificial intelligence more accessible.

The Gemma 3 models are available in various sizes, offering developers options ranging from one billion, four billion, 12 billion and 27 billion parameters. This flexibility enables AI engineers to select the optimal model based on hardware and performance requirements. For instance, larger and more complex models can be utilized on GPUs or tensor processing units, while smaller models are better suited for smartphones.

Gemma is built on the same technical research as Google’s Gemini model, the company’s most advanced and powerful AI to date. Gemini powers the Gemini AI chatbot, formerly known as Bard, which is accessible on the web and mobile devices and supports many of Google’s AI-driven services.

According to Google, Gemma 3’s technical design enables it to deliver high performance relative to its size, surpassing larger models like Llama-405B, DeepSeek-V3, and OpenAI’s o3-mini in preliminary human preference evaluations on the LMArena leaderboard.

Even when running on a single device or GPU, Gemma delivers sufficient power for developers to build AI applications with multimodal capabilities, including advanced text and visual reasoning. It features a 128,000-token context window, allowing it to process approximately 30 high-resolution images, a 300-page book, or over an hour of video—comparable to the context length of OpenAI’s GPT-4o.

Google stated that the Gemma 3 model family supports function-calling and tool-use capabilities, allowing developers to automate tasks and create AI agents. With its large context window, Gemma 3 can process extensive data and streamline complex sequential tasks.

Alongside Gemma 3, Google introduced ShieldGemma 2, a 4-billion-parameter variant designed to assess image safety and categorize them as safe or harmful.

ShieldGemma allows developers to create applications that analyze uploaded images for potentially harmful content. It generates safety labels across three categories: “dangerous content,” “sexually explicit,” and “violence.”

Developers integrating ShieldGemma into their applications can customize the model by specifying content to monitor and label. With open-source weights and parameters, it can be further trained to meet various industry-specific requirements and controls.