Google Unveils Gemma 3: Advanced AI Models for Single GPU/TPU

Google has introduced Gemma 3, a collection of lightweight, state-of-the-art open models designed to run efficiently on a single GPU or TPU, according to a company blog post. These models are built from the same research and technology that powers Google's Gemini 2.0 models and are available in various sizes, including 1B, 4B, 12B, and 27B, to cater to different hardware and performance needs.

Gemma 3 models offer several new capabilities, such as support for over 140 languages, advanced text and visual reasoning, and a 128k-token context window for handling complex tasks. Additionally, they introduce quantized versions to reduce model size and computational requirements while maintaining high accuracy.

Alongside Gemma 3, Google has launched ShieldGemma 2, a 4B image safety checker that provides safety labels across three categories: dangerous content, sexually explicit, and violence. This tool is built on the Gemma 3 foundation and is designed to promote responsible AI development.

Gemma 3 integrates seamlessly with popular tools like Hugging Face Transformers, JAX, Keras, and PyTorch, allowing developers to customize and deploy models across various platforms, including Google Cloud TPUs and NVIDIA GPUs. The models are also featured on the NVIDIA API Catalog, enabling rapid prototyping with an API call.