TwelveLabs

TwelveLabs builds foundation models that understand video: models that search, describe and answer questions about footage rather than generate it. Based in San Francisco and Seoul and led by co-founder and CEO Jae Lee, it sells two model lines and an agent built on them. Marengo is a multimodal embedding model for search. Pegasus is a video-language model that writes text about what it sees. Jockey, launched publicly on October 15, 2026, is an agent that combines the two. Its customers are in media and entertainment, sports, advertising, security, automotive and the public sector, and since October 2026 it has also pitched its models for training robots. It's the best-known specialist in video understanding, a field otherwise dominated by general multimodal models such as Gemini.

Models (October 2026)

Separating retrieval from description is TwelveLabs' core design choice. An embedding index lets semantic search run across a large archive cheaply, and the language model is applied only to the clips that matter. Marengo's embeddings can be stored in standard vector databases, so the approach fits existing retrieval stacks.

Video for Physical AI

Pegasus 1.6 is TwelveLabs' first model built to understand egocentric video: first-person footage of someone cooking, assembling parts or operating a robot remotely. The company aims it at physical AI, turning such footage into training data for robots. It supports five workflows: action segmentation and labeling, dense captions describing spatial relationships, quality scoring to filter weak clips, natural-language search for rare events, and flagging of faces, bystanders and sensitive data for consent and compliance review. It's one of the clearer commercial bets on learning from human video (see imitation learning).

Funding and Cloud

TwelveLabs raised a $50 million Series A in June 2024, co-led by NEA and NVIDIA's NVentures. On July 1, 2026 it raised a $100 million Series B, co-led by NEA and NAVER Ventures, with Amazon, Radical Ventures, Korea Investment Partners, Index Ventures, Quadrille Capital and Red Bull Ventures. As part of that round, AWS became its preferred cloud: new models launch there first, and the companies are tuning video inference for AWS Trainium chips. TwelveLabs' models have been available fully managed in Amazon Bedrock since July 15, 2025, starting with Marengo 2.7 and Pegasus 1.2.

How It Compares

General multimodal models, Gemini above all, can watch video directly and are often good enough for one-off questions about a clip. TwelveLabs' advantage is at archive scale: embeddings that make thousands of hours searchable, a model tuned for video rather than images, and enterprise distribution through Bedrock. Open vision-language models and on-device tools for computer vision compete at the low end, especially for privacy-sensitive footage. Many editing pipelines skip video models entirely and have an agent read extracted frames and contact sheets, which is cheaper but loses motion and sound. For the trade-offs, see video understanding; for where it feeds editing, see agentic video editing.