19/03/2025
Google has released Gemma 3, a new suite of open-source AI models. These models are built on the same technology as Gemini 2.0 and are designed to be efficient and accessible.
Gemma 3 is available in different sizes (1B, 4B, 12B, and 27B parameters) to suit various hardware needs, from mobile devices to powerful workstations. The models support multiple types of data, including text, images, and short videos (excluding the 1B model).
The models are pre-trained in over 140 languages, with native support for over 35 languages, enabling global application development.
Gemma 3 provides performance comparable to larger models, specifically in tasks requiring reasoning and problem-solving. It is designed to run on single GPUs or TPUs, reducing reliance on cloud resources.
The models offer an extended context window of up to 128k tokens and support function calling for task automation.
Google also introduced ShieldGemma 2, a 4B model for image safety, which categorizes images based on safety levels.
Gemma 3 is available on platforms like Hugging Face and integrates with tools such as Llama and Google’s AI Studio. It can be deployed on Vertex AI, Cloud Run, the Google GenAI API, or locally.