Twelve Labs

Twelve Labs develops video foundation models and an accompanying API/SDK platform for what it calls video intelligence: ingesting raw video and turning it into indexed, searchable, AI-ready data. Its two core models are Marengo, a multimodal embedding model that analyzes frames and their temporal relationships together with speech and sound for search and any-to-any retrieval, and Pegasus, a video-first language model that combines visual, audio and speech signals to generate text such as summaries, chapters and highlights. The platform exposes Search, Analyze and Embed products, plus an agentic offering the company markets as Jockey, described as a video intelligence AI agent.

Core workflows advertised include natural-language search across video libraries (actions, scenes, dialogue and emotions without manual tagging), automatic content segmentation and scene-break detection, compliance and brand-safety review, highlight and rough-cut assembly that exports into editing workflows, and aggregate analysis of video for editorial or creative insight. The company states its pipeline ingests multimodal data at roughly 60x real time, indexing an hour of video in about a minute and more than 10,000 hours per day, and cites SOC 2 Type II certification with deployment options in the customer's chosen environment. Developer resources include official Python and JavaScript SDKs, an MCP integration, a Claude Code plugin, sample applications and integrations such as an Embed API pairing with MongoDB Atlas Vector Search.

External profiling describes the company as specialized multimodal video-understanding infrastructure with proprietary video foundation model architectures, including a gated modality experts design and temporal modeling intended to treat video as a continuous event rather than a sequence of static frames.

Founding story

Co-founder and CEO Jae Lee has described the company as starting roughly five years before mid-2026 from the observation that the world does not happen in text but in motion, and that most of the world's video remains inaccessible to machines except through filenames, folders, captions, transcripts and human memory. Prior to founding Twelve Labs he was lead data scientist at the Cyber Operations Command of the Republic of Korea Ministry of National Defense, with earlier software engineering experience at Amazon and Samsung Electronics.

Business model

Twelve Labs sells access to its video foundation models through a usage-based API and SDK platform, supplemented by enterprise contracts and cloud-marketplace distribution via partners such as AWS Bedrock.

Tiered usage-based pricing: a free plan covering up to 10 hours of indexing with limits on index duration, volume and concurrency; a pay-as-you-go Developer plan priced per minute of video indexing and analysis, per 1,000 search queries and per 1,000 output tokens; and an Enterprise plan on committed-use contracts with custom pricing.

Traction

Third-party profiling puts team size at 70-80 and lists partnerships with NVIDIA, Oracle, AWS, Intel, Blackmagic Design, Vidispine, EMAM and Blackbird. The company reports platform throughput of more than 10,000 hours of video indexed per day and publishes benchmark claims including a 13.1% advantage for Pegasus 1.5 over Gemini 3.1 Pro on multimodal prompting and a top ranking on Video-MME.

Latest developments

On 1 July 2026 the company announced a $100 million Series B co-led by NEA and NAVER Ventures with participation from Amazon, Radical Ventures, Index Ventures, Korea Investment Partners, Quadrille Capital and Red Bull Ventures. Alongside the round, AWS signed a multiyear contract to host Twelve Labs workloads on Trainium chips, with new models debuting on the service for developers.