Google has unveiled Gemma 3, a portable AI model for researchers

Google is continuing to evolve its Gemma lineup aimed at researchers and developers. The new version, Gemma 3, offers improved portability and can even run on a single GPU, making it much more accessible to users working with AI outside of cloud servers.

The new version, Gemma 3, offers improved portability and can even run on a single GPU, making it much more accessible to users working with AI outside of cloud servers.

Key Improvements to Gemma 3

In contrast to previous versions, Gemma 3 has received major improvements in performance and flexibility of use. It supports 35 languages and the pre-learned models cover more than 140. The context window has increased to 128,000 tokens, allowing AI models to analyze longer and more complex queries.

Google offers four versions of the model: 1B, 4B, 12B and 27B parameters. All but the youngest version can handle not only text but also images.

Optimize and run on a single GPU

The main feature of Gemma 3 was its portability. Typically, running AI models requires powerful compute clusters with multiple GPUs. However, Gemma 3 can even run on a single GPU or tensor processing unit (TPU), which significantly lowers the entry threshold for researchers.

This level of optimization Google has achieved through quantization, a technique that reduces model size and computational resource requirements without sacrificing accuracy. For maximum performance, the company recommends using NVIDIA GPUs ranging from the Jetson Nano to top-of-the-line Blackwell.

Additional Tools and Security

For developer convenience, Gemma 3 is integrated with Vertex AI and Google Colab, allowing the model to be customized for specific tasks. Google says the new version outperforms competitors including Llama 405B, DeepSeek-V3 and o3-mini.

Besides Gemma 3, the company introduced ShieldGemma 2, a tool for checking the security of images. This AI model (4B parameters) analyzes content and flags dangerous, explicit or violent images, ensuring more responsible use of generative AI.

An AI model (4B parameters) analyzes content and flags dangerous, explicit or violent images, ensuring more responsible use of generative AI.

Accessibility and Outlook

Researchers and developers can download Gemma 3 via Kaggle and Hugging Face, and get more information from Google AI Studio. Its improved portability and reduced hardware requirements may make it one of the most popular models among independent AI researchers.