Google Unveils Gemma 3: A New Open-Source AI Model

Google has announced the release of its latest open-source AI model, Gemma 3, which is designed for optimal performance on mobile devices. This new iteration comes in four distinct sizes, identified by the number of parameters it contains: 1 billion, 4 billion, 12 billion, and 27 billion.

In addition to the Gemma 3 model, Google has also made available ShieldGemma 2, a specialized security model that integrates with applications to enhance safety measures.

According to Google, Gemma 3 surpasses several larger models, including Meta's Llama 405B and Deepseeks V3, in certain benchmarks. However, such comparisons may not fully reflect the practical applications of these models. Notably, Gemma 3 is pre-trained in 140 languages and can utilize an additional 35 languages without prior training.

The capabilities of Gemma 3 have been improved to better process images, text, and short video clips compared to its predecessor. The model's context window has been expanded to accommodate up to 128,000 tokens, allowing for more extensive inputs. Furthermore, Gemma 3 supports 'Function Calling,' which enables the model to leverage various tools, facilitating the automation of tasks.

Additionally, Google has included quantized versions of Gemma 3, which reduce the model's size and computational requirements while maintaining high accuracy.

As for ShieldGemma 2, it is equipped with 4 billion parameters and is designed to flag images that may raise security concerns. The model's weights are open, enabling users to customize it for their specific needs. ShieldGemma can be integrated into applications to alert users when content that may depict inappropriate themes, such as violence or sexual imagery, is generated.

Both models are accessible through popular platforms like HuggingFace, Google AI Studio, and Kaggle. Additionally, Google has established a separate program for academics, allowing selected individuals to gain access to Google Cloud Credits.

In the so-called 'Gemmaverse,' users can connect and share experiences related to these models, fostering a community of innovation and collaboration.