The company introduced Gemini, its largest and most capable AI model, marking the next step in their journey toward making AI helpful for everyone. Constructed from the ground up to be multimodal, Gemini demonstrates the ability to generalize and seamlessly understand, operate across, and combine different types of information, encompassing text, images, audio, video, and code. This endows Gemini with sophisticated multimodal reasoning and advanced coding capabilities. The model comes in three different sizes — Ultra, Pro, and Nano — providing flexibility to run on everything from data centers to mobile devices. The training of Gemini at scale took place on the company’s AI-optimized infrastructure, utilizing Google’s Tensor Processing Units (TPUs) v4 and v5e. Concurrently, the company also unveiled its most powerful and scalable TPU system to date, Cloud TPU v5p.
Gemini is now accessible in some of the company’s core products: Bard employs a fine-tuned version of Gemini Pro for advanced reasoning, planning, understanding, and more. Pixel 8 Pro, the inaugural smartphone engineered for Gemini Nano, utilizes it in features such as Summarize in Recorder and Smart Reply in Gboard. Initial experiments with Gemini in Search have resulted in accelerated Search Generative Experience (SGE). Early in the upcoming year, Gemini Ultra will be introduced to enhance the Bard Advanced experience. Over the coming months, Gemini is slated to power features in additional products and services like Ads, Chrome, and Duet AI.
For Android developers interested in building Gemini-powered apps on-device, an early preview of Gemini Nano, the most efficient model, is now available through Android AICore. Starting December 13, developers and enterprise customers can access Gemini Pro via the Gemini API in Vertex AI or Google AI Studio, a free web-based developer tool. The company plans to refine Gemini Ultra, including comprehensive trust and safety checks, and will initially make it available to select groups before a broader release to developers and enterprise customers early next year.
SOURCE: Google