Google must have taken us before we offer Jima 3. In fact, while JEMA 1 was released in February 2024 and Jema 2 arrived in May this year, the Mountain View will wait more before lifting the curtains on the third repetition of the giant model. Particularly the wait announcement, since the popularity of these models is very high, which has recorded more than 100 million downloads a year. At the same time, the Gem Sever Environmental System has more than 60,000 variations of the model flourish. Remember that Jema is open source, while Gemini owns it.
Designed to perform directly to various equipment from smartphone to work station, Gema 3 offers complimentary flexibility for developers who want to connect AI without needing heavy infrastructure.
Improved on a single Exceller
Google has claimed the title “Best Model on the Best Model on the same Excelor in the world”, saying it has crossed references to Limrina Diagnosis 405 B, Deep Sak-V3 and O-3-mini.
This model is based on the same technical bases such as Gemini 2.0 Series and several sizes: 1B, 4B, 12B and 27b. Enough to cover widespread use according to the developers’ requirements.
It has become a habit, with the beginning of each new model, it is a benchmark. © Google
Cut for multi -modal and multi -linguistic
One of the strongest points of JEMA 3 is in its modern reasoning capabilities, whether the text or the image. A progress has been made up to 128,000 tokens through a context window to treat audio dialogues and long documents more effectively. The model is also capable of analyzing short photos and video streams.
The linguistic component is not to start: JEMA takes care of 35 languages and relies on training in 140 languages, thus facilitating it internationally.
Developers have not been forgotten
Google has also integrated several tools to exploit the model more easily, for example an option with an option “function calling and structural output” that makes it possible to design more efficient intelligent agents. Google has also promised a proper security device with Shieldjima 2, a 4B image control module that attributes three standards labels: sensitive, clear or violent material.
The development process has also been monitored through strict data governance and risk assessment policies, especially in scientific and technical fields.
Is now available
Developers can test JEMA 3 on Google AI Studio today. And for those who prefer, the model can also be downloaded from a cogl or a hug.
