Google Introduces Gemma 3: The Leading Single-Accelerator Model

In early February 2024, Google launched version 1 of Gemma, which was followed by version 2 in May. Now, the tech giant has introduced Gemma 3 as the latest open model aimed at developers.

Gemma is Google’s family of open models that have remarkably achieved over 100 million downloads in just one year. This family includes around 60,000 variants in a space referred to as the Gemmaverse. The models are designed to operate quickly on various devices, including smartphones, laptops, and high-end workstations.

Gemma 3 harnesses the same research and technology that powers the Gemini 2.0 models. It is available in multiple sizes: 1 billion, 4 billion, 12 billion, and 27 billion parameters.

Google claims that Gemma 3 stands out as the “world’s best single-accelerator model,” meaning it outperforms its competitors when running on a single GPU or TPU. In tests conducted on LMArena, it excelled over other significant models such as Llama-405B, DeepSeek-V3, and o3-mini.

One of the notable features highlighted by Google is the advanced text and visual reasoning capabilities available in the 4B+ model versions. This includes the ability to analyze images, text, and short videos effectively. Gemma 3 supports an impressive 128k-token context window and is compatible with over 35 languages right out of the box, with pre-trained support for more than 140 languages.

• AI-Powered Workflows: Gemma 3 allows for the creation of AI-driven workflows using function calling, enabling users to automate tasks and create interactive experiences.
• Efficient Performance: The model offers improved performance thanks to quantized models, which reduce the size and computational demands while maintaining precision.

On the safety front, Google has introduced a robust image safety checker called ShieldGemma 2, designed to categorize content across three critical safety categories: dangerous content, sexually explicit material, and violence. The development of Gemma 3 also emphasized data governance aligned with Google’s stringent safety policies, involving detailed fine-tuning and rigorous benchmark evaluations.

Gemma 3’s enhanced capabilities in STEM-related tasks have led to a thorough assessment regarding the potential for misuse in creating unsafe content; results indicate that the risk is low.

Users can experience Gemma 3 firsthand on Google AI Studio. Additionally, model downloads can be found on Kaggle and Hugging Face.