Google Lanza Gemma 3n, AI model open with capabilities and high performance on local devices

The new model is designed for low Gamma devices, including mobile equipment, offering large capacities and performance in large number of tasks.

***

• Gemma 3n It is the new open model of Google With support for image, audio, video and text.
  
• It is designed to run on devices with only 2GB of memory thanks to its efficient architecture.
  
• It is already available on platforms such as Hugging Face, Kaggy and Google Ai Studio.

Google He officially presented Gemma 3n, The new generation of its family of open artificial intelligence models, standing out for its multimodal approach and its local execution capacity with limited resources, including mobile devices.

After a previous view during the event Google I/O, The complete model is now available for developers to download and use it freely. Unlike Gemini, that is closed and focused on mass consumption, Gemma is designed for independent development and researchreview NEWIN.

What is Gemma?

Gemma It is the line of open models of Google, different from its owner platform Gemini Its objective is to provide accessible and modifiable tools for developers and researchers. With the version 3n, The company introduces native support for image, audio, video and text entries, which represents a significant leap with respect to its previous versions based only on text.

The incorporation of these multimodal capacities allows generating text from different types of data, facilitating their integration into mobile applications, educational tools, intelligent assistants and more.

One of the most prominent advances of Gemma 3n is its base architecture, called MatFORM. According to Google, This design allows to contain smaller functional versions within a larger architecture, such as a Russian doll. In this way, a single model can operate in different sizes according to the type of task, optimizing the use of resources.

The two main sizes available are E2B and E4b, with 5,000 million and 8,000 million parameters, respectively. However, thanks to innovations such as Per Layer Embeddings (PLE) and new coders for audio and vision, its memory use is maintained equivalent to models of 2,000 and 4,000 million parameters. This allows the model to work even on devices with only 2GB of RAM.

Improved performance and capabilities

Google states that Gemma 3n It offers substantial improvements in tasks of reasoning, coding and multilingualism. It supports 140 languages ​​for text processing, and 35 languages ​​in its multimodal understanding.

In the computer vision section, the new encoder is used Mobilenet-V5, Designed to function efficiently even on mobile phones. This component is capable of processing 60 fps video on devices such as Pixel of Google.

For its part, the audio encoder allows you to perform tasks in voice recognition and translation directly in the device, without the need for cloud connection.

Possibilities of immediate use

Interested developers can access Gemma 3n immediately through platforms such as Hugging Face, Kaggy, and Google Ai Studio. This early availability opens the door for rapid adoption in AI projects that require local execution, either for privacy reasons, energy efficiency or cost.

Besides, Gemma 3n It is positioned as the first model with less than 10,000 million parameters to exceed the 1,300 points in the test LMARENA, A meter for the general quality of language models.

With the launch of Gemma 3n, Google It is strongly positioned in the accessible and efficient the segment, responding to both the demands of the developer community and the technical needs of the Edge Computing.

The possibility of having a powerful, versatile and functional model in limited hardware represents a unique opportunity to create more independent, private and personalized tools. This marks an important step towards a more distributed artificial intelligence ecosystem, without depending completely on cloud solutions.

Article written by a content editor. Edited by Angel Di Matteo / Diariobitcoin

Original image of Diariobitcoin, created with artificial intelligence, freely used, licensed under public domain

WARNING: Diariobitcoin offers informative and educational content on various topics, including cryptocurrencies, AI, technology and regulations. We do not provide financial advice. Cryptactive investments are high risk and may not be adequate for all. Investigate, consult an expert and verify the applicable legislation before investing. I could lose all its capital.

Subscribe to our newsletter