Google presents Gemma 3, its new “most capable” open model that can be executed in a single GPU

Gemma is the family of open source AI models developed by Technological, which was initially presented in February last year: 2,000 million parameters (2b) and 7,000 million parameters (7b), and which updated in May in the framework of its annual developer conference with the launch of Gemma 2, with which it reached a size of 27,000 million parameters The height of flame 3 of finish.

Now, Google has expanded the family with its new gemma 3 model, to which he defines as “the most capable model” to execute in a single graphic processing unit (GPU)available in sizes of 1, 4, 12 and 27 million parameters to choose the option that best suits the specific needs of ‘hardware’ and performance.

So, These models are designed to run quickly on any device, from ‘smartphones’ to laptops or work stationsso they have been devised with the aim of helping developers to create applications of AI “wherever they need them.”

In addition, it is a collection of open and light models that are based on “the same research and technology” that drives its Gemini 2.0 models, as Google has detailed in a statement in its blog.

In this sense, the company has stressed that Gemma 3 offers high performance for its sizes, which ensures that it surpasses other models of the sector such as Llama-405b of Meta, Depseek-V3 and O3-MINI of OpenAI, in preliminary evaluations in the classification of Lmarena. Therefore, it allows to create “attractive user experiences” that adapt to a single GPU or TPU host, specifically, with optimized performance in the NVIDIA GPUs.

Following this line, the new open model of AI is compatible with 140 languages, and offers immediate support for more than 35 languages, which allows developers to create applications in the language of their users.

On the other hand, Google has indicated that with Gemma 3 developers can also easily create applications that analyze short images, text and videos. It also offers a 128K Tokens context window to allow applications to process and understand large amounts of information.

Another option that Gemma 3 allows the calls for functions and structured output, which enables task automation, as well as creating agent experiences in applications.

Google has also presented Shieldgemma 2, its new image safety verifier with 4B size based on Gemma 3. This solution evaluates the images and generates security labels in three categories, such as dangerous, sexually explicit and violence content.

In this sense, developers can use Shieldgemma 2 to integrate it into their applications in order to customize their safety needs for users. It also takes advantage of Gemma 3’s performance and architecture to promote the responsible development of AI.

In addition to all this, Google stressed that, when developing the model, a “thorough” risk assessment has been carried out based on its security protocols, including alignment with its security policies through “precise adjustments and solid comparative evaluations”.

With all this, Gemma 3 and Shieldgemma 2 are already available to integrate into workflows with tools such as Hugging Face, Kaggy, Google ai Edge or VerTex AI, among other options. You can also access Gemma 3 capabilities from Google Ai Studio.