Google Unveils Gemma 3: Lightweight AI for Single GPU
Google has launched Gemma 3, the latest addition to its lightweight open model series, designed to run efficiently on phones, laptops, and other devices.
In a March 12 blog post, Google stated that Gemma 3 is built on the same research and technology as its Gemini 2.0 models, ensuring smooth user experiences on a single GPU or TPU host.
These models support text and visual inputs but can only generate text outputs. With parameter sizes of 1 billion, 4 billion, 12 billion, and 27 billion, they provide developers with options to suit various AI applications.
According to the Gemma 3 page on Hugging Face, the 27B model was trained on 14 trillion tokens, the 12B model on 12 trillion tokens, the 4B model on 4 trillion tokens, and the 1B model on 2 trillion tokens.
Although Gemma 3 was trained on a text-based dataset, Google has not disclosed specific data sources. However, the company confirmed that Gemma 3’s model weights are open-source, allowing developers to create pre-trained and instruction-tuned variants of the small language model (SLM).
Additionally, Gemma 3 features a 128k-token context window, enabling it to process and understand larger volumes of information.
Gemma 3 vs. Other AI Models
Google claims that Gemma 3 outperforms Meta’s Llama-405B, OpenAI’s o3-mini, and DeepSeek-V3 in preliminary human preference evaluations conducted on LMArena, a crowdsourced AI benchmarking platform developed by UC Berkeley researchers.
Gemma 3: Capabilities and Deployment
Gemma 3 enables developers to create AI applications capable of analyzing images, text, and short videos. It supports over 35 languages for linguistic tasks and has pre-trained capabilities for more than 140 languages. Additionally, developers can leverage Gemma 3 to build AI-powered automation tools and agent-based applications, thanks to its structured outputs and function-calling support.
The model is available for download on Kaggle and Hugging Face and can also be accessed via Google Studio.
Google highlighted multiple deployment options for Gemma 3, including Vertex AI, Cloud Run, Google GenAI API, local environments, and other platforms, allowing users to select the most suitable infrastructure for their applications.
Developers can further train and fine-tune these models on Google Colab, Vertex AI, and even on gaming GPUs. Google also introduced a revamped codebase with optimized fine-tuning and inference recipes to enhance efficiency.
Also read: OnePlus swaps Alert Slider for Smart Button.
What is ShieldGemma 2?
Alongside Gemma 3, Google introduced ShieldGemma 2, an AI safety tool with four billion parameters.
ShieldGemma 2 is designed to tag AI-generated images with labels such as dangerous content, sexually explicit material, and violence. Google stated that it can be integrated with various developer tools and offers customization options for enhanced adaptability.