Google introduces Gemma 3, the new generation of open source AI models that run on a single GPU

Gemma 3 represents a significant advance in the accessibility of artificial intelligence, offering high-performance open source models that can run on a single GPU or TPU. Developed by Google, this new generation of open models is available in four sizes (1B, 4B, 12B, and 27B) to adapt to different hardware and performance needs.

Among its most notable capabilities is support for over 140 languages, allowing the creation of truly global applications. Additionally, it incorporates advanced visual and textual reasoning capabilities, being able to analyze images, texts, and short videos, opening new possibilities for interactive applications.

The new family of models features an expanded context window of 128,000 tokens, allowing it to process large amounts of information, and supports function calls to automate tasks and build experiences with AI agents.

Along with Gemma 3, Google has launched ShieldGemma 2, a 4B image safety checker that provides safety labels in three categories: dangerous content, sexually explicit, and violence.

According to preliminary evaluations, Gemma 3 outperforms models like Llama3-405B and DeepSeek-V3, positioning itself as the best open source AI model that can run on a single accelerator. This represents an important step toward democratizing AI technology, allowing developers worldwide to create advanced applications without the need for costly infrastructure.