Google DeepMind has unveiled the latest iteration of its generative AI models—Gemma 3. This significant update introduces multi-modal capabilities that enable the models to process and understand visual data alongside text. With Gemma 3, users can now analyze images, identify objects, and answer questions about the content within images, marking a major step forward in AI’s ability to understand and interpret complex visual information.
Launched on March 12, Gemma 3 is available for experimentation in the Google AI Studio, providing developers with a powerful platform for AI development. This new version also brings substantial improvements to the model’s abilities in mathematics, coding, and following instructions, expanding its utility across various domains. Notably, Gemma 3 supports vision-language inputs and text outputs, and can process up to 128k tokens within its context window, making it capable of handling large, intricate datasets and complex tasks with ease.
Gemma 3 is available in four different sizes—1B, 4B, 12B, and 27B parameters—giving developers flexibility depending on their specific needs. These models are also available in pre-trained and general-purpose instruction-tuned versions, allowing for a range of applications. The update supports over 140 languages and introduces enhanced features such as structured outputs and function calling, improving its capabilities for reasoning and interaction. This makes it a highly versatile tool for developers across different industries.
To help developers integrate and deploy Gemma 3, Google offers several options including Cloud Run and the Google GenAI API. Additionally, the model’s code base has been revamped for better performance, with pre-built recipes available for inference and fine-tuning. For those who wish to experiment with or further develop the model, Gemma 3 model weights are available for download via Kaggle and Hugging Face. To ensure maximum performance, Nvidia provides direct support for the Gemma 3 models on a range of GPUs, and Google Cloud TPUs are optimized for use with this model.