Google Deepmind introduced Gemma 3, an update for family generative AI models with multimodality that allows models to analyze images, answer images, identify objects, and perform other tasks that include analysis and understanding of visual data.
The update was announced on March 12 and can be tested in the Google AI Studio for Development. Gemma 3 also significantly improves mathematical, coding and instructions according to Google Deepmind’s abilities.
Gemma 3 supports visual language inputs and text outputs, processes context windows up to 128,000 tokens and understands more than 140 languages. Improvements were also made for mathematics, reasoning and chat, including structural outputs and function calls. Gemma 3 comes in four sizes of “friendly developers” 1b, 4b, 12B and 27B and in pre -school and generally paid and tuned versions. “The 128K-Cen context window allows Gemma 3 to process and understand the massive amacent information, easily solve complex tasks,” said Google Deepmind’s.