Google Releases Gemma 3, the Most Powerful AI Model to Run on a Single GPU

Google has released an open source artificial intelligence model, Gemma 3, a new member of a family of models that developers can use to create applications that can run locally on workstations or even smartphones. It supports 35 languages, and has text, image, and short video analysis capabilities.

Image source: blog.google

Google described Gemma 3 as the “world’s best single-GPU model,” and assured that it demonstrates results better than competitors from Meta✴, DeepSeek, and OpenAI when running on a single GPU; it is optimized for Nvidia hardware and other AI accelerators. The video analysis component has been updated to support high-resolution recordings and non-square frame shapes; the ShieldGemma 2 input and output image security filter responds to images of a sensitive nature and materials containing violence.

Google has an open AI model of the third generation, but the potential of this area was helped to reveal by the Chinese DeepSeek – its projects demonstrated that there is a demand for AI solutions with relatively low system requirements. Gemma 3 has significant knowledge in the field of technical disciplines, and Google conducted additional testing for the possibility of misuse of the model, for example, to create harmful substances – the probability of failure of protective mechanisms in the company was assessed as low.

The open license of Gemma 3 is in fact limited: Google has specified what it can be used for. The new model is available, among other things, in Google Cloud. As part of the academic program, you can apply for $10,000 in credits if you use it for research purposes.