Google Unveils New Line of AI Models Gemma 4 Based on Gemini 3 Technology

Google has officially launched its latest series of artificial intelligence models, known as Gemma 4, which are developed using the advanced Gemini 3 technology. This new collection comprises four distinct variants designed for compatibility with a range of devices, making them versatile for developers and users alike. The models are distributed under the Apache 2.0 open license, which provides developers with extensive freedom to use and adapt the models according to their specific needs.

The Gemma 4 family consists of four models that vary in the number of parameters, a critical factor that significantly influences the model's performance. For smartphones and other compact devices, Google has crafted versions with 2 and 4 billion parameters. Meanwhile, for more powerful computers, models featuring 26 billion parameters (a type known as 'mixture of experts') and 31 billion parameters have been developed, catering to a broader range of computational capabilities.

Google has highlighted its achievement in reaching an impressive level of efficiency, managing to extract more 'intelligence' per parameter. According to the company, the models with 31 and 26 billion parameters secured the third and sixth positions in the Arena AI text ranking, outperforming systems that possess 20 times more parameters. This remarkable performance underscores significant progress in AI development, showcasing the ability to execute complex tasks with high precision.

All models in the Gemma 4 series are capable of processing images and videos, including the ability to recognize text within them. The smaller versions also support audio processing and speech recognition. Notably, these models can generate code without needing an internet connection, enhancing their usability in environments with limited network access. Furthermore, the models have been trained on over 140 languages, making them universally applicable across various linguistic contexts.

The Gemma 4 models are released under the Apache 2.0 license, which empowers developers to have complete control over their data, infrastructure, and the models themselves. This flexibility allows for seamless deployment of the models in any environment, whether locally or in the cloud. Such an approach ensures adaptability of the technology to meet the specific requirements of users.

Developers can experiment with the new models through popular platforms such as Hugging Face, Kaggle, and Ollama. This accessibility enables developers and researchers to explore the new capabilities offered by the Gemma 4 models and incorporate them into their projects. In this way, Google continues to advance its artificial intelligence solutions, providing new tools for the creation of innovative products and services.