Titan Multimodal Embeddings G1

This model takes in plain text of about 8.1 pages at a time and gives back a numeric code for it, not written words. A buyer sits down with it to sort a large pile of articles or reports by subject, or to match a question to the right passage in a long file. Unlike the picture-matching model in its family, it takes in no photos or captions and is built for whole long documents. It stops where a job needs a written reply, since it only produces codes that computers compare, so that kind of work needs a different kind of model.

This model takes a piece of text of up to about eight pages and turns it into a numeric code that computers can compare, so it finds and groups similar writing rather than writing anything itself. Sit down with it to sort thousands of documents by topic, power a search that matches a question to the right passage, or flag near-duplicate text. Unlike its family's picture-matching sibling, which reads only a caption's worth of text, it digests full long documents in one go. It never composes answers or prose, so any job that needs a written reply needs a different kind of model.

Amazon Nova Micro 1.0 takes text and gives back text, and it is the quickest to respond in the Nova family, reading up to about 128 pages. Send it text to summarise, translate, sort or brainstorm over, or a quick chat turn; it handles simple maths and basic code. Replies stop at about 5 pages.

Nova 2 Lite reads text, images, video and files, answers in text, can think step by step before replying, and keeps about 1,000 pages in mind at once. Give it documents to pull facts from, video to extract details from, code to write, or a multi-step task that uses the tools you connect to it. Nova Lite 1.0 reads only text and images, about 300 pages, and Nova Premier handles the hardest reasoning.