Google has announced the early preview release of Gemma 3n, an open-source artificial intelligence model designed for efficient, on-device performance. Revealed at Google I/O 2025, Gemma 3n is available for developers to experiment with today. Built in partnership with Qualcomm, MediaTek, and Samsung, Gemma 3n features a new architecture optimized for mobile devices, enabling fast, multimodal AI experiences on phones, tablets, and laptops. The model is based on Gemini Nano technology and can be deployed on platforms such as Kaggle. Gemma 3n supports audio, text, image, and video inputs, and is capable of running with as little as 2GB or 3GB of RAM. It is available in configurations with a 4 billion (4B) and 2 billion (2B) parameter active memory footprint, with the ability to dynamically switch between them using MatFormer-based 'nested' architecture. The model supports 140 languages, can operate on CPUs using LiteRT, and achieves high Chatbot Arena Elo scores. Innovations such as Per-Layer Embeddings reduce memory usage, and the model offers advanced multilingual and multimodal capabilities, including speech recognition and translation. Developers can access Gemma 3n via Google AI Studio and Google AI Edge, and it is expected to be integrated into Android, Chrome, and other open-source libraries. Separately, the United Arab Emirates' Advanced Technology Research Council has released Falcon Arabic and Falcon H1, new AI models aimed at advancing Arabic-language artificial intelligence. Falcon Arabic is now live on Hugging Face and is designed to capture the linguistic diversity of the Arab world, while Falcon H1 uses a hybrid Transformer-Mamba architecture to optimize performance and portability, reportedly outperforming models from Meta and Alibaba.