Gemini 1.0
Google released a family of models in three sizes, trained on text, images, audio and video at once rather than assembled from separate parts.
Why it matters
Multimodality stopped being an add-on: the model trained on every kind of data from the beginning.
The three variants, Ultra, Pro and Nano, target complex tasks, general use and on-device work respectively. The company reported passing the human level on the MMLU benchmark. The demonstration video was later found to be edited in a way that overstated the capability, which damaged trust in the announcement. The model brought together the Google Brain and DeepMind teams, merged in April 2023.