Google introduces the Gemma 3 family of accessible lightweight models

Continuing a campaign to make its artificial intelligence models more accessible, Google LLC today announced the next generation of its lightweight open source family. Gym Large language models that can run on the same graphics processing unit.

New Jema 3 models come A range of Size Allow The developers will choose between 1 billion, 4 billion, 12 billion and 27 billion parameters. These boundaries allow AI engineers and developers to select the best models for hardware and performance requirements. For example, if it would be Are running GPU or Tenser on Processing Unit – Allows Large More Complex models – or smartphone – require a small model.

Gemma’s primary technology shares technical research, like Google’s gymnasium model, is a very complex and powerful model developed by the company to date. Gemini is power Gemini Ai Chat BoatBard named before, Which is Available on web and mobile devices and is also used to provide many Google -based services.

With this technical design, Google said that Jima 3 is capable of providing its size LIGH high performance, which improves large models such as Lama-405B, Dipic-V3 and Openi Open Meni in the early human preferential diagnosis on the Lamrina Leader Board.

Even when running on a single device or GPU, Jema still provides developers with advanced text and visual reasoning with multi -modal capabilities to create AI applications. It has a 128,000 token context window, which means it can handle close resolution images, a 300 -page book or an hour or more video. These Openi’s GPT4OO is compared to the length of context.

Google said the model family includes the functioning of the function calling and tolls that Equivalent Enable developers to automatically make tasks and build AI agents. Combined with a large context window, JEMA 3 Will be able Eat large amounts of data and automatically make complicated order tasks.

Along with JEMA 3, Google also announced Shieldjima 2, which is a different form of 4 billion parameters that can check the images of safety and make them safe or dangerous.

Shieldjima developers enables to make applications Who can check Pictures uploaded uploads of potentially hazardous content. This is out of safety labels In three different types Like “Dangerous Material,” “sexually Clarity“And” violence. “

It can further prepare the model for their needs by providing the developers who use it in their applications and label the content. Its weight and parameters are also open source So It can be trained Different needs and control of the industry.

Pictures: RashGoogle