Google Unveils Gemma 3, a Lightweight AI Model Designed for Devices

Google’s new AI model, Gemma 3, is designed for efficiency on devices ranging from smartphones to laptops. Offering multiple parameter sizes and a host of features, it paves the way for sophisticated applications across various languages. Alongside it, ShieldGemma 2 aims to improve AI safety, showcasing Google’s commitment to responsible AI innovation.

Google has unveiled Gemma 3, the latest addition to its lineup of lightweight AI models aimed at running on devices like smartphones and laptops. Launched on March 12, 2025, this new series harnesses the same foundational technology that drives the Gemini 2.0 models. Google emphasized its efficiency, stating, “This helps you to create engaging user experiences that can fit on a single GPU or TPU host.”

These models are versatile, designed to process both text and visual data but, interestingly, they can only generate text outputs. Developers can choose from various model sizes, including one billion, four billion, 12 billion, and 27 billion parameters, depending on their application needs. Each version is tailored for specific performance capabilities; for instance, the 27B model has been trained using an impressive 14 trillion tokens.

While Google trained Gemma 3 on a vast and varied text dataset, it hasn’t disclosed specific sources of the data. A notable feature is that its weights are open-source, enabling developers to create tailored pre-trained variants as well as instruction-tuned versions of this small language model (SLM). Another exciting feature is a 128k-token context window that allows it to process extensive amounts of information.

In comparisons to existing AI models, Google claims Gemma 3 has outperformed Meta’s Llama-405B and OpenAI’s o3-mini in early evaluations on LMArena, a benchmarking platform by UC Berkeley. Besides, the model offers robust capabilities, allowing for analyses of images and text, accommodating over 35 languages, and pretrained support for more than 140 languages. It provides developers with tools to automate tasks and capabilities leveraged by AI agents, thanks to its structured outputs.

For those interested, Gemma 3 is readily available for download on platforms such as Kaggle and Hugging Face, as well as through Google Studio. The company stated, “Gemma 3 offers multiple deployment options, including Vertex AI, Cloud Run, the Google GenAI API,” and more, giving flexibility across infrastructures.

Furthermore, developers can continue to refine the models through platforms like Google Colab, Vertex AI, or even on gaming GPUs. Google also included a new optimized codebase in Gemma 3, which comes with efficient recipes for fine-tuning and inference.

In addition to Gemma 3, Google announced ShieldGemma 2, a safety tool designed for AI. This four billion parameter model allows developers to label AI-generated content such as potentially dangerous or explicit material. Google assures that it can easily integrate with other tools, providing options for customization to enhance safety in AI-generated outputs.

With the introduction of Gemma 3, Google is making strides in the lightweight AI model space, offering flexibility and power for developers working across various devices. Its performance claims against other leading models and features, like a vast context window and support for multiple languages, are particularly noteworthy. Moreover, alongside ShieldGemma 2, Google shows its commitment to enhancing the safety of AI-generated content, a growing concern in this technology’s evolution.

Original Source: indianexpress.com