TwelveLabs

Search works across spoken words, sounds, and visuals at once, so you can find a scene by describing an action, a line, or a mood. No tags or labels are required ahead of time. The Marengo model returns ranked matches with an uncertainty signal on each one, which helps you judge how much to trust a result.

Marengo turns video, audio, images, text, and documents into one shared embedding space. That lets you run any-to-any retrieval, such as matching a text query to a video clip or an image to a document. Version 3.5 produces 512-dimensional embeddings and a sync endpoint that accepts up to 16K tokens in a single query.

Pegasus reasons over the full timeline of a clip rather than sampling a few frames. It tracks entities, cause and effect, and narrative across time, then writes summaries, captions, chapter markers, or answers to direct questions. The company reports it outperforming Gemini 3.1 Pro on multimodal prompting by 13.1% in its own tests.

The platform ingests video at roughly 60 times real-time speed, so an hour of footage indexes in about a minute, and it can handle more than 10,000 hours per day on the paid tiers. Faster indexing means fresh footage becomes searchable sooner, which matters for newsrooms and live production. AI video intelligence only pays off when the results arrive while they still matter.

TwelveLabs ships a REST API plus Python and Node SDKs, and an MCP connection for AI assistants. Teams can call search, embed, and analyze endpoints directly, or plug the platform into existing editing and asset tools instead of building retrieval from scratch.

The platform carries SOC 2 Type II certification and encrypts data in handling. It also supports private or self-managed deployment for organizations that cannot send footage to a shared cloud, which is often the deciding factor for government and regulated buyers.