Google DeepMind has introduced a new collection of generative models called Gemma 3, available for use on a variety of devices, including smartphones, laptops, and workstations. This collection includes lightweight models that offer high speed and performance, making them ideal for use on a single GPU or TPU. Gemma 3 supports over 140 languages and provides advanced text and visual analysis capabilities, opening up new possibilities for creating interactive and intelligent applications.
Among the new features of Gemma 3 is the ability to handle complex tasks thanks to an expanded context window of 128,000 tokens and support for function calling to automate processes. Gemma 3 models are also available in quantized versions, ensuring faster performance and reduced computational requirements. This allows users to adapt the models to their needs using various platforms and tools such as Hugging Face Transformers, PyTorch, and others.
Together with Gemma 3, Google DeepMind has launched ShieldGemma 2—an image safety verification tool that provides automatic content labeling across three safety categories: dangerous, sexually explicit, and violent. This allows developers to customize the tool according to their needs, ensuring the safe development of AI solutions. ShieldGemma 2 is built on Gemma 3, giving it high efficiency and flexibility.
Gemma 3 easily integrates into existing workflows, offering a variety of deployment options, including Vertex AI, Cloud Run, and local environments. This gives users the ability to quickly start working with the models, adapt them, and deploy them on different hardware platforms. Gemma 3 is optimized for use on NVIDIA GPUs and integrates with AMD GPUs via the ROCm™ stack, making it accessible to a wide range of users and developers.
