Alphabet, the parent company of Google, has released its new collection of open-source Artificial Intelligence (AI) models. Gemma 3, available in sizes of 1, 4, 12, and 27 million parameters, is the company’s most powerful model capable of running on a single graphics processing unit (GPU), with the aim of helping developers create AI applications on the go.
Gemma is the family of open-source AI models developed by the California-based tech company, which was initially presented in February of last year in two sizes: 2 billion parameters (2B) and 7 billion parameters (7B). The model was updated in May within the framework of its annual developer conference with the launch of Gemma 2, which reached a size of 27 billion parameters (27B) and, with it, performance on par with Meta’s Llama 3. Moreover, Google has disclosed that Gemma has reached the milestone of 100 million downloads.
Open-source AI on any device
The models are designed to run quickly on any device, from smartphones to laptops or workstations, so they have been designed with the aim of helping developers create AI applications “wherever people need them.” Additionally, it is a collection of open and lightweight models that are based on “the same research and technology” that powers its Gemini 2.0 models, the business explains.
In this regard, the company has emphasised that Gemma 3 offers high performance for its sizes, ensuring that it surpasses other industry models such as Meta’s Llama-405B, DeepSeek-V3 (which has also partially open-sourced its model), and OpenAI’s o3-mini in preliminary evaluations in the LMArena ranking. Therefore, it allows the creation of “engaging user experiences” that adapt to a single GPU or TPU host, specifically, with optimised performance on Nvidia GPUs.
Multilingual and multinational AI
Following this approach, the new open AI model is compatible with 140 languages and offers immediate support for more than 35 languages, allowing developers to create applications in the language of their users. On the other hand, Google has pointed out that with Gemma 3, developers will also be able to easily create applications that analyse images, text, and short videos. Likewise, it offers a context window of 128K tokens to allow applications to process and understand large amounts of information. Another option that Gemma 3 enables is function calls and structured output, allowing for task automation and creating agent experiences in applications.
ShieldGemma also gets an update
Google has also introduced ShieldGemma 2, its new 4B-size image security verifier based on Gemma 3. This solution evaluates images and generates security labels in three categories: dangerous content, violent content, and sexually explicit material. In this regard, developers can use ShieldGemma 2 to integrate it into their applications to customise their security needs for users. Likewise, it leverages the performance and architecture of Gemma 3 to promote the responsible development of AI.
With all this, Gemma 3 and ShieldGemma 2 are now available to be integrated into workflows with tools such as Hugging Face, Kaggle, Google AI Edge, or Vertex AI, among other options. The capabilities of Gemma 3 can also be accessed from Google AI Studio, the company explains.
