Dec 2023
Google Announces Gemini 1.0 — Natively Multimodal Model Family
Google announced the Gemini 1.0 model family in three sizes — Ultra, Pro, and Nano — built to be natively multimodal from the ground up, understanding and generating across text, images, audio, and video in a single unified model. Gemini Ultra reportedly surpassed GPT-4 on 30 of 32 benchmarks evaluated, including the MMLU academic knowledge test. Gemini Nano was designed to run on-device on Pixel 8 Pro smartphones without a network connection.
Why it mattersNatively multimodal architecture — rather than bolted-on vision — proved that training modalities jointly from the start produced stronger cross-modal reasoning than adapter-based approaches.
Try itIn retrospect, the pattern to notice is that Gemini Nano's on-device capability established that sub-2B models fine-tuned for a narrow task can match larger cloud models on that task.