Traditional AI tools can't watch your video library and tell you the exact moment someone smiles in frame 47 of file 2,304. Twelve Labs built something that can.
This is video search that actually works at scale. You feed it natural language queries—it returns precise timestamps across speech, visuals, audio, and on-screen text. Two models work together here: Marengo handles spatial and temporal reasoning. Pegasus processes video-language understanding. They've beaten benchmarks from major cloud providers (matters when you're dealing with petabytes of footage).
Developers get three main APIs. Search finds specific moments in massive libraries. Analyze generates text summaries or reports through prompting. Embed creates multimodal vector embeddings for recommendation engines. A media company could ask "show me all scenes where products appear on kitchen counters"—then get back exact clips from thousands of hours of content.
The free tier gives you under 10 hours of indexing in a shared environment. Good for testing. After that you'll need to move up. Options include cloud, private cloud, or on-premise deployment. Custom model training on your own data is available too.
The shared environment limitation on the free plan means you can't really stress-test performance at production scale. But for building proof-of-concept applications? It's enough to see whether the video understanding actually delivers. The $12 million in seed extension funding suggests other developers think it does.
