# TwelveLabs > Enterprise video AI platform delivering multimodal video understanding through two production foundation models — Marengo (multimodal embedding) and Pegasus (video-language) — accessible via REST API, Python SDK, and Node.js SDK. Founded in 2021; headquartered at 55 Green Street, San Francisco, CA 94111 with R&D operations in Seoul, South Korea. Models are designed natively for video (vision + audio + speech + text) rather than adapted from image or text models, and are deployable in shared cloud, dedicated cloud, private VPC, and on-premise configurations. Available natively on AWS (Amazon Bedrock + Marketplace), Snowflake (Cortex integration), and Databricks (Mosaic AI). Investors include NVentures (NVIDIA's venture capital arm), Databricks Ventures, Snowflake Ventures, SK Telecom, HubSpot Ventures, and IQT; advisory board includes Fei-Fei Li and Silvio Savarese (Stanford), Jeffrey Katzenberg (former DreamWorks Animation CEO), Alex Wang (Scale AI CEO), Lukas Biewald (Weights & Biases CEO), Nicolas Dessaigne (Algolia founder), and Jay Simons (Atlassian president). SOC 2 Type II certified. --- ## Platform - [Platform Overview](https://www.twelvelabs.io/product/product-overview): Describes the end-to-end video intelligence platform: index ingest → multimodal foundation models (Marengo for embeddings, Pegasus for generation) → three API surfaces (Search, Analyze, Embed). Supports Python and Node.js SDKs. Free tier limited to 10 hours of indexing across up to 100 videos per index with 5 concurrent indexing tasks; Developer tier is pay-as-you-go with 100,000 videos per index and 25 concurrent tasks; Enterprise tier scales custom. The platform indexes visual, audio, speech, and on-screen text in a unified pipeline. Documentation lives at docs.twelvelabs.io. - [Video Search API](https://www.twelvelabs.io/product/video-search): Semantic search API that returns time-stamped, confidence-scored moments rather than whole videos. Accepts text, image, audio, and composed multimodal queries. Indexes four modalities simultaneously — visual content, audio, speech (ASR), and on-screen text — across 100+ languages. Common deployments: production scene selection, archive activation, customer-facing video search, surveillance investigation, and contextual ad placement. Integration pattern: instantiate client → create index → upload via task → query semantically. Domain-specific fine-tuning available for vertical accuracy gains. - [Analyze API](https://www.twelvelabs.io/product/analyze): Pegasus 1.5-powered analysis API for video-to-text generation. Outputs include summaries, chapters with timestamps, Q&A pairs, highlights, hashtags, topic categories, time-stamped event descriptions, and free-form text per natural language prompt. Supports videos up to 2 hours per request. Processing is asynchronous (`task.wait_for_done()` pattern). Outputs are structured JSON with non-overlapping temporal segments containing start/end timestamps and metadata fields populated against a user-defined schema. Used for social media post generation, highlight reels, chapter generation for long-form content, and interview/assessment scoring. - [Embed API](https://www.twelvelabs.io/product/embed): Marengo-retrieval-2.7 (and Marengo 3.0) embedding generation API. Returns segment-level 512-dimensional float embeddings with metadata (scope, time offsets) for video, image, text, and audio in a unified vector space. Marengo 3.0 embeddings are 6x more compact than Amazon Nova's 3072d and 3x more compact than Google Vertex's 1408d while outperforming both. Designed for RAG pipelines, recommender systems, contextual ad personalization, content moderation, anomaly detection (e.g., corrupt video identification), and training-data quality curation. Bring-your-own vector database — no proprietary lock-in. --- ## Models - [Models Overview](https://www.twelvelabs.io/product/models-overview): Side-by-side comparison of the two foundation models. **Marengo** (current version 3.0; prior versions 2.6, 2.7) is the multimodal video encoder — converts video, image, text, audio, and speech into a shared 512-dimensional embedding space for any-to-any retrieval, semantic search, hybrid search, and anomaly detection. **Pegasus** (current version 1.5; prior versions 1, 1.2) is the video-language model — performs schema-driven segmentation, structured metadata generation, summarization, chaptering, Q&A, and open-ended text generation conditioned on video. Now supplemented by dedicated marketing pages for each model (below); this overview page remains the direct side-by-side comparison. - [Marengo](https://www.twelvelabs.io/marengo): Dedicated product page for the Marengo video embedding model, positioned as "find anything, in any video." Headline specs: 4 hours of continuous video and audio processed with full temporal context in a single call; 36+ languages for query and retrieval, native rather than translated; 512-dimensional embeddings, 6x smaller than competitors; 30x faster than competitors at video indexing while leading on accuracy. Named capabilities not offered by general-purpose embedding models: composed queries mixing image, text, and audio into a single embedding (e.g., a player photo plus a text description); unified 36-language search space with no translation step or cross-language accuracy loss; search by example frame (image query refined by text). Page-level comparison table sets Marengo 3.0 against Gemini Embedding 2 and Amazon Nova Multimodal Embeddings on embedding dimensions (512 vs 3072 default/1536/768 via Matryoshka vs 3072 default/1024/384/256 via Matryoshka), max video length (4 hrs continuous vs 120 sec/request vs 30 sec/segment chunked), multimodal composite queries (single embedding for image+text+audio vs no native composite query API for either competitor), sports recognition (5 sports vs not a claimed capability for either competitor), latency in seconds per second of video (0.05 vs ~0.50 vs 1.50+), and embedding confidence scores (exposed vs not exposed vs not exposed). States 73% composite video-retrieval performance vs. Google Vertex (52%) and Amazon Nova (55%), matching the same figure in the Marengo 3.0 blog post below. Page promotes an MCP Server ("Connect Claude, Cursor, or any MCP client to your video index") alongside Quick Start, Sample Apps, and SDK/API onboarding paths — all four surfaced as a "signup to first result in 5 minutes" flow. - [Pegasus](https://www.twelvelabs.io/pegasus): Dedicated product page for the Pegasus video-to-text model, positioned as "turn raw video into structured data." Headline specs: 2 hours max video duration with full temporal context retained end-to-end; 12 languages supported across input prompts and generated output; 0 pre-indexing steps (send a URL, asset, or base64, get text back); JSON-native structured segmentation output. Named capabilities: exact-timestamp answers rather than approximate location; reads on-screen text, jersey numbers, whiteboards, and receipts alongside the speech track; converts a raw video into structured JSON against user-defined segment types (speaker change, brand appearance, scene cut) and fields; accepts a reference image (logo, face, product) as visual context for a query. Page-level comparison table sets Pegasus 1.5 against Gemini 3.1 Pro and GPT-5.5 on max single-call duration (120 min vs 90 min vs not specified), structured segmentation output (JSON-native, schema-conditioned vs generic Structured Outputs with no native temporal segmentation for both competitors), multimodal prompting, and per-definition time windows (native to Pegasus, general-purpose only for the other two). This comparison set (Gemini 3.1 Pro, GPT-5.5) differs from the Gemini-only benchmark comparison in the Pegasus 1.5 blog post below. Page promotes the same MCP Server / Quick Start / Sample Apps / SDK onboarding flow as the Marengo page. --- ## Pricing & Enterprise - [Pricing](https://www.twelvelabs.io/pricing): Three tiers with published per-unit rates. **Free**: 0 cost; 10 hours (600 min) monthly indexing; 100 videos/index; 5 concurrent tasks; 90-day index retention. **Developer (pay-as-you-go)**: Video indexing $0.042/min; infrastructure $0.0015/min; Search API $4 per 1,000 queries; Embed API $0.042/min (video), $0.0083/min (audio), $0.10 per 1,000 image requests, $0.07 per 1,000 text requests; Analyze API $0.0292/min (input video) + $0.0075 per 1,000 output tokens. Unlimited video hours; 10,000 hours/index; 100,000 videos/index; 25 concurrent tasks. **Enterprise**: Custom contracts, capacity scales with monthly spend. - [Pricing Calculator](https://www.twelvelabs.io/pricing-calculator): Interactive estimator that converts video hours indexed, monthly search queries, embed calls (by modality), and analyze API input minutes + output tokens into a monthly cost projection at Developer-tier rates. Useful for benchmarking against alternatives or for budgeting before contracting to Enterprise. - [Enterprise](https://www.twelvelabs.io/enterprise): Enterprise tier features dedicated capacity, custom fine-tuning, cloud / private cloud / on-premise deployment, encrypted data handling, and advanced governance. Targets sports, advertising and marketing, security, media and entertainment, and government. Publicly named enterprise customers and partners include NFL Media (Brad Boim, Senior Director), Sejong City (CTO-level engagement), MindsDB (Jorge Torres, Co-Founder and CEO), Voxel51 (Daniel Gural, ML Evangelist), Mindprober (Pedro Almeida, CEO), Source Digital (Michael Philips, CPO), Protege, AffiliateNetwork, and Dyn Sport. Cited outcomes: 10x faster content review and compliance scanning; single-video processing completed in 4 hours through one API call. --- ## Solutions - [Media and Entertainment](https://www.twelvelabs.io/solutions/media-and-entertainment): Broadcasters, studios, and streaming platforms apply the platform to archive segmentation (decades of footage → scene-level searchable intelligence), production acceleration (raw footage → rough cut), content repackaging (single source → trailers, social clips, regional edits), tag-free metadata generation, duplicate detection, and personalized discovery. Designed for petabyte-scale archives. The MLSE case study reports a 97% reduction in content search time (16 hours → 9 minutes) and manual logging cut from 30+ hours to seconds. - [Advertising](https://www.twelvelabs.io/solutions/advertising): Use cases include contextual ad placement (matching ads to scene-level theme, tone, and emotion), real-time brand-safety detection, creative intelligence (identifying messaging and moments that drive engagement), accelerated production through automated rough cuts, semantic ad-inventory discovery, and hyper-contextual personalization. Sits inside the broader CTV measurement workflow for ad platforms and agencies. - [Government](https://www.twelvelabs.io/solutions/government): Natural-language search across millions of video hours sourced from CCTV, drone, and satellite feeds for public-sector and investigative workflows. Capabilities include incident summarization in minutes, recurring-pattern and anomaly detection, cross-source intelligence fusion, time-stamped confidence-scored results, and automated incident documentation. Deployment options span cloud, private cloud, and on-premise, including air-gapped configurations; fine-tuning on customer-specific data is available for domain accuracy. Previously combined with Security under a single "Government and Security" page; the two are now separate Solutions entries covering the same underlying capability set. - [Security](https://www.twelvelabs.io/solutions/security): Surveillance review and security-operations applications, split out from the former combined Government and Security page into its own Solutions entry. Covers incident detection, recurring-pattern and anomaly identification across CCTV and camera feeds, and rapid retrieval of specific moments across large surveillance archives, with the same cloud / private cloud / on-premise / air-gapped deployment flexibility and fine-tuning options as the Government solution. - [Sports and Broadcasting](https://www.twelvelabs.io/solutions/sports-and-broadcasting): Highlight generation, live and archived footage search, and broadcast workflow automation for sports content. New Solutions entry that extends the media-and-entertainment use cases (archive segmentation, tag-free metadata, rapid rough cuts) into sports-specific workflows such as live-event highlight packages and multi-camera archive search. Draws on outcomes already demonstrated in the MLSE and Dyn Sport case studies. --- ## Developers - [Developer Hub](https://www.twelvelabs.io/developer-hub): Central entry point for developers — API reference, Python SDK, Node.js SDK, code samples, tutorials, and integration guides for the Search, Analyze, and Embed APIs. Connects to the Playground (auth.twelvelabs.io/u/login) for no-code experimentation, and to documentation at docs.twelvelabs.io. - [Sample Apps](https://www.twelvelabs.io/sample-apps): Reference implementations and starter projects across GitHub. Covers semantic video search applications, RAG pipelines that ground LLMs in video content, automated highlight generators, content classifiers, and multimodal embedding workflows. Each sample is paired with a tutorial. - [Research](https://www.twelvelabs.io/research): Research team holds 100+ publications across top AI/CV/NLP conferences and 5+ wins in global competitions. Research areas: video-native foundation models, multimodal alignment, perceptual reasoning, video-language pretraining, and multi-vector embeddings. Published benchmark TWLV-I provides holistic evaluation methodology for video foundation models. Co-organized inaugural NeurIPS 2024 Workshop on Video-Language Models. - [Whitepapers](https://www.twelvelabs.io/whitepapers): In-depth technical and business documents covering multimodal video AI architecture, enterprise deployment patterns, evaluation methodology, vertical use cases, and ROI analyses for media, advertising, and security buyers. --- ## Case Studies - [Case Studies Index](https://www.twelvelabs.io/case-studies): Aggregated library of customer deployments spanning sports, media, commerce, nonprofit, government, and security verticals with quantified outcomes. - [Qencode](https://www.twelvelabs.io/case-studies/qencode): Full-stack video platform (transcoding, live streaming, storage, CDN, player, and analytics via unified APIs) based in Los Angeles. Qencode's existing ML-driven encoding already reduces file sizes by up to 60% at matched quality, but the platform remained "semantically opaque" — encoded video with no understanding of its content. Qencode embedded TwelveLabs Marengo and Pegasus directly into its encoding pipeline so video is encoded and made fully searchable in a single API call, rather than running video intelligence as a separate downstream process. Teams now search encoded libraries in plain language ("highlights of the latest live performance," "any scene with a visible logo") and get back exact clips with timestamps. Use cases span media/broadcasting archives, e-commerce product video, education/enterprise training libraries, and security/compliance monitoring. Quoted executive: Murad Mordukhay, CEO of Qencode — "They were the only provider whose models truly understood video the way we always envisioned." - [GS SHOP](https://www.twelvelabs.io/case-studies/gs-shop): South Korea's leading omnichannel commerce platform (TV live channels, data home shopping, mobile live commerce; part of GS Retail). Its "Short Pick" experience presents product highlights in ~30-second clips. GS SHOP's existing recommendation engine treated video as a passive thumbnail — a customer watching a running-shoe clip would just see more running shoes, with no read on styling or context. GS SHOP layered TwelveLabs Marengo (via Amazon Bedrock) onto the existing engine, building a four-axis "Appeal Point" framework (functionality, style and sentiment, use situation, practicality) so video context — not just product tags — drives cross-category recommendations. Validated via a Korean-language proof-of-concept query ("a shopping host placing a piece of kimchi on boiled pork and eating it with chopsticks") that Marengo located in seconds from a raw one-hour broadcast tape. Results: +57.5% lift in ordering customers, +29.4% conversion rate lift, +21.7% more unique clicks, average watch time up from 6.3s to 8.0s, and broadcast video search time cut from 1–2 hours to seconds. Quoted executive: Lee Chang-min, Part Lead, GS SHOP Search & Recommendation Part. - [UNICEF Korea](https://www.twelvelabs.io/case-studies/unicef-korea): UNICEF Korea Committee holds 8TB+ of media accumulated across decades of fundraising campaigns and child rights programs — video and images spanning relief activities, donor events, and nonprofit operations across dozens of countries, scattered across individual PCs and network storage with no unified search. Manual cataloging, keyword search, and frame-by-frame AI tools all failed at this scale. TwelveLabs — implemented by Korean AI integration partner Letsur — migrated the archive to AWS S3 and indexed approximately 200 hours / 2TB of video plus associated images and documents using Marengo (multimodal embeddings) and Pegasus (event localization across full video duration), enabling natural-language search such as "children collecting water at a field site in Africa." New field content is now indexed automatically with no manual classification step. Impact: 95% reduction in content retrieval time, 8TB+ of fragmented data transformed into a structured searchable library, and thousands of manual folder reviews eliminated from campaign planning and donor communications workflows. Quoted: Jae Lee, TwelveLabs CEO (not a UNICEF Korea spokesperson) — "UNICEF Korea's media archive is more than a storage repository — it's a first-hand record of where support is needed and how programs operate in the field." - [MLSE (Maple Leaf Sports & Entertainment)](https://www.twelvelabs.io/case-studies/mlse): MLSE owns eight professional teams including the Toronto Maple Leafs (NHL), Toronto Raptors (NBA), Toronto FC (MLS), Toronto Argonauts (CFL), Toronto Marlies (AHL), Raptors 905 (NBA G League), TFC II (MLS Next Pro), and Raptors Uprising Gaming Club (NBA 2K League). Deployed semantic video search across the production archive. Outcomes: content search time reduced 97%; video search and retrieval cut from 16 hours to 9 minutes; manual logging cut from 30+ hours to seconds; content discovery compressed from days to minutes. - [SBS](https://www.twelvelabs.io/case-studies/sbs): SBS is one of South Korea's leading television and radio broadcasters, spanning dramas, sports, news, and reality TV. Deployed TwelveLabs multimodal AI alongside SBS's proprietary scene-search technology to enable reuse of archived media assets and scene-level search across internal and individual-team archives — addressing slow retrieval of visual effects, unaccounted-for metadata from individually uploaded footage, and time-consuming footage trading and sharing. Broadcast media deployment covering semantic content search and management at network scale. - [Dyn Sport](https://www.twelvelabs.io/case-studies/dyn-sport): Dyn Media is a sports-focused streaming platform covering 3,000+ live events per season. Deployed Marengo 2.7 for video indexing and Pegasus 1 for automated descriptions, summaries, and captions across their content library. Result: highlight packages generated within minutes of events; activation of previously underutilized archive footage; reduced manual search dependency. Moritz Michael Mücke, Head of Product & Tech at Dyn Media, is the quoted executive. - [AffiliateNetwork](https://www.twelvelabs.io/case-studies/affiliatenetwork): Video analysis applied to affiliate marketing content for performance measurement and creative intelligence. - [Protege](https://www.twelvelabs.io/case-studies/protege): Video understanding for talent development and coaching use cases. --- ## Partners - [Partners Overview](https://www.twelvelabs.io/partners): Index of strategic technology partners (NVIDIA, AWS, Databricks, Snowflake), channel partners, media-vertical integrators (e.g., Monks), and tool integrations (e.g., Avid Media Composer panel). - [NVIDIA](https://www.twelvelabs.io/partners/nvidia): Strategic investor and infrastructure partner. TwelveLabs SaaS platform runs on NVIDIA GPU acceleration on AWS, with models optimized for NVIDIA GPU architectures for lower latency and higher throughput. NVIDIA AI Blueprint for Video Search & Summarization (VSS) is available on GitHub for joint customer deployments. Joint engineering focuses on production video intelligence performance. - [AWS](https://www.twelvelabs.io/partners/aws): Available natively in Amazon Bedrock for managed video understanding, and in AWS Marketplace for direct procurement and deployment. Tutorial content covers S3 Vectors integration and trailer/highlight generation workflows. Joint go-to-market focuses on media workflow modernization. Partnership content refreshed June 2026. - [Databricks](https://www.twelvelabs.io/partners/databricks): Databricks Ventures has invested in TwelveLabs. Integration enables generating multimodal video embeddings via the TwelveLabs Embed API, storing them in Delta tables, and running vector search workflows at scale through Mosaic AI Vector Search. Roadmap includes native model access through Mosaic AI Model Serving so developers can call Pegasus and Marengo directly inside the Databricks Data Intelligence Platform. - [Snowflake](https://www.twelvelabs.io/partners/snowflake): Snowflake Ventures has invested in TwelveLabs. Integration with Snowflake Cortex allows customers to connect TwelveLabs to existing Snowflake environments for semantic video search, video-grounded RAG, and content recommendations without moving data out of the warehouse. Joint quickstart and webinar resources cover sports storytelling and AI video search. - [Monks](https://www.twelvelabs.io/partners/monks): Media and advertising channel partner. - [Avid Media Composer Integration](https://www.twelvelabs.io/integrations/twelvelabs-panel-for-avid-media-composer): TwelveLabs panel for Avid Media Composer brings semantic video search directly into the editing workflow used across broadcast and post-production. - [Become a Partner](https://www.twelvelabs.io/partners/become-a-partner): Partner program entry point for technology, channel, and integration partners. --- ## Resources - [Blog](https://www.twelvelabs.io/blog): Technical blog covering model releases, research, customer stories, and engineering deep dives. Posts authored by named team members including Kian Kim and Shannon Hong (model release leads). - [Introducing Pegasus 1.5 (June 1, 2026)](https://www.twelvelabs.io/blog/introducing-pegasus-1-5): Release announcement for Pegasus 1.5, a step beyond the clip-based Q&A behavior of 1.2. Pegasus 1.5 performs schema-driven segmentation — partitioning entire videos according to a user-defined schema and labeling each non-overlapping temporal segment with structured metadata. Handles up to 2 hours of video in a single request. Outputs structured JSON with start/end timestamps. On TwelveLabs' internal evaluation set, Pegasus 1.5 scores 0.4279 on segmentation vs Gemini 3.1 Pro at 0.3370, and 0.4555 on multimodal prompting vs Gemini 3.1 Pro at 0.3243 (a 13.1-percentage-point lead). Authored by Kian Kim and Shannon Hong with contributions across Science, ML Engineering, Data, and Product. - [Marengo 3.0 (November 30, 2025)](https://www.twelvelabs.io/blog/marengo-3-0): Release announcement for Marengo 3.0. Achieves 73% composite performance (mean of MSRVTT, MSVD, Vatex, YooCook2, SomethingSomething-v2, Kinetics-700, DiDemo, Visual7W-Pointing, QVHighlight, PasskeyRetrieval, HagridRetrieval, WikipediaRetrieval, Urban1K, Dense-WebVid-CoVR, and several sports/object benchmarks) while maintaining the fastest normalized latency at 0.05 seconds per second of video, vs Google Vertex at 52% composite (~0.5 sec/sec latency) and Amazon Nova at 55% composite (>1.5 sec/sec latency, making it 30x slower than Marengo 3.0). 512-dimensional embeddings — 6x more compact than Amazon Nova (3,072d) and 3x more compact than Google Vertex (1,408d). The 73% composite breaks down into four modality-level composites: video retrieval 70.2% (mean of 14 benchmarks incl. MSRVTT, MSVD, VATEX, YooCook2, DiDemo, QVHighlight, Dense-WebVid-CoVR, and 7 sports benchmarks; a 5.8-point lead over Marengo 2.7, 25 points over Vertex, 18.2 over Nova), audio retrieval 73.2% (mean of GTZAN, Clotho, Librispeech), image retrieval 92.2% (mean of Visual7W-Pointing, Urban1K, TextCaps, Object365-medium, OpenLogo), and text retrieval 88.3% (mean of SomethingSomethingv2-MC, Kinetics-700, PasskeyRetrieval, HagridRetrieval, WikipediaRetrieval). Individual benchmark scores: MSRVTT 72.5% (vs Marengo 2.7 at 71.7%, Vertex 59.5%, Nova 69.6%); MSVD 78.2%; VATEX 86.4%; Librispeech audio retrieval 99.7% (vs Nova <10%, Vertex unsupported); GTZAN music genre 75.5% and Clotho audio captioning 44.2% (vs Marengo 2.7 71.9%/41.4%, Nova 64.9%/29.5%); SoccerNet-Action 79.4 mAP (vs Nova 23.0, Vertex 21.5); PasskeyRetrieval 94.4% NDCG@1 (vs Gemini Embedding text-only 38.5%, Nova 82.5%); Wikipedia Multilingual 85.6% NDCG@10 (vs Marengo 2.7 29.0%); TextCaps OCR 89.2%; OpenLogo 79.2% mAP; Object365-medium 99.1%. Multilingual: Korean 87.2%, Japanese 91.1% on Crossmodal 3600. New capabilities: composed retrieval (visual + text queries) and hour-long video processing. Text retrieval improved 58% → 88.3% vs Marengo 2.7. Authored by Dan K., Jeff K., Chris J., Jeremy K., Cooper H., Royce H. - [Press](https://www.twelvelabs.io/press): Press releases, funding announcements, executive interviews, and analyst recognition. CB Insights AI 100 honoree three consecutive years; coverage in Bloomberg and major AI/business outlets; NVIDIA investment announced. - [Explore](https://www.twelvelabs.io/explore): Curated showcase of platform capabilities through demos and interactive examples. --- ## Webinars & Events - [Webinars](https://www.twelvelabs.io/webinars): Recorded sessions with technology partners (Quickplay, Mimir, Adobe) and category education for media, archive, and advertising teams. - [NAB Show 2026](https://www.twelvelabs.io/nab-show-26): TwelveLabs presence at the National Association of Broadcasters 2026 conference for media buyers. - [Broadcast Asia 2026](https://www.twelvelabs.io/broadcast-asia-26): TwelveLabs presence at Broadcast Asia 2026 for APAC media customers. - [AWS re:Invent](https://www.twelvelabs.io/aws-reinvent): Joint AWS re:Invent presence covering Bedrock and Marketplace deployments. --- ## Company - [About Us](https://www.twelvelabs.io/about-us): Founded in 2021 by twelve founding members spanning language, video, machine learning, and perception research. Mission: pioneer multimodal, video-native AI that sees and understands like humans do. Headquarters: 55 Green Street, San Francisco, CA 94111. Additional office: Seoul, South Korea. Advisory board: Fei-Fei Li (Stanford), Silvio Savarese (Stanford), Jeffrey Katzenberg (former DreamWorks Animation CEO), Alex Wang (Scale AI), Lukas Biewald (Weights & Biases), Nicolas Dessaigne (Algolia), Jay Simons (Atlassian). Focus areas include memory layers for video intelligence, air-gapped deployment, and time-based metadata pipelines. - [Careers](https://www.twelvelabs.io/careers): Open roles across research, ML engineering, platform engineering, go-to-market, and operations. Offices in San Francisco and Seoul. - [Contact](https://www.twelvelabs.io/contact): Sales, partnerships, and general inquiry channels. Sales engagement is the primary entry point for Enterprise tier conversations. - [Security](https://www.twelvelabs.io/security): Security and compliance posture. Encryption: TLS 1.2 or higher for data in transit; AES-256 minimum for data at rest. Access controls follow identity and least-privilege with regular access reviews and role-based permissions. Defense-in-depth architecture. AWS underlies infrastructure. Compliance reports and detailed sub-processor information available via the Trust Center at trust.twelvelabs.io. Security contact: security@twelvelabs.io. SOC 2 Type II certified per the Enterprise page. - [Patents](https://www.twelvelabs.io/legal/patents): Intellectual property notice covering TwelveLabs patents and patent applications. --- ## Key Facts - **Founded:** 2021 by twelve founding researchers covering language, video, ML, and perception. CEO and co-founder: Jae Lee. - **Headquarters:** 55 Green Street, San Francisco, CA 94111. R&D office: Seoul, South Korea. - **Production foundation models:** Marengo 3.0 (multimodal embeddings; 512d; released November 30, 2025) and Pegasus 1.5 (video-language; up to 2-hour single-request video; released June 1, 2026). - **Marengo 3.0 headline benchmarks:** 73% composite video-retrieval performance vs. Google Vertex (52%) and Amazon Nova (55%), consistent across the Marengo product page and the Marengo 3.0 blog post. MSRVTT 72.5%; SoccerNet-Action 79.4 mAP (vs Nova 23.0 / Vertex 21.5); PasskeyRetrieval 94.4% NDCG@1 (vs Gemini 38.5%); Librispeech audio retrieval 99.7%; multilingual support including Korean (87.2%) and Japanese (91.1%) on Crossmodal 3600. - **MCP Server:** Available for connecting Claude, Cursor, or any MCP client directly to a video index; promoted on both the Marengo and Pegasus product pages as part of the standard onboarding flow alongside Quick Start, Sample Apps, and SDKs. - **Pegasus 1.5 headline benchmarks:** 0.4279 segmentation quality vs Gemini 3.1 Pro 0.3370; 0.4555 multimodal prompting quality vs Gemini 3.1 Pro 0.3243 (13.1% lead on both) - **Scale claims:** Platform processes 10,000+ hours of video per day across customer workloads. - **API surfaces:** Search, Analyze, Embed — accessible via REST API, Python SDK, and Node.js SDK; documentation at docs.twelvelabs.io. - **Deployment options:** Shared cloud, dedicated cloud, private VPC, on-premise, air-gapped (government). - **Cloud and data platform distribution:** AWS (Bedrock + Marketplace), Snowflake (Cortex + Snowflake Ventures investor), Databricks (Mosaic AI + Databricks Ventures investor). - **Investors and partners:** NVentures (NVIDIA's venture capital arm, strategic + GPU infrastructure), Databricks Ventures, Snowflake Ventures, SK Telecom, HubSpot Ventures, IQT. - **Advisory board:** Fei-Fei Li, Silvio Savarese, Jeffrey Katzenberg, Alex Wang, Lukas Biewald, Nicolas Dessaigne, Jay Simons. - **Security certifications:** SOC 2 Type II. TLS 1.2+ in transit; AES-256 at rest. Trust Center at trust.twelvelabs.io. - **Research footprint:** 100+ publications in top AI conferences; 5+ wins in global competitions; co-organized inaugural NeurIPS 2024 Workshop on Video-Language Models; published TWLV-I evaluation benchmark. - **Industry recognition:** CB Insights AI 100 honoree three consecutive years. - **Solutions verticals:** Media and Entertainment, Advertising, Government, Security, Sports and Broadcasting. - **Named customers and partners (public):** NFL Media, MLSE, Dyn Media, SBS, Sejong City, MindsDB, Voxel51, Mindprober, Source Digital, Protege, AffiliateNetwork, Qencode, GS SHOP, UNICEF Korea Committee. --- ## Legal - [Privacy Policy](https://www.twelvelabs.io/legal/privacy-policy): Data collection, processing, retention, and user rights under GDPR and CCPA. - [Terms of Use](https://www.twelvelabs.io/legal/terms-of-use): Service terms, acceptable use, customer obligations. - [Terms of Service](https://www.twelvelabs.io/legal/tos): Master service terms for paid plans. - [Acceptable Use Policy](https://www.twelvelabs.io/legal/aup): Acceptable use constraints across the platform. - [Data Processing Addendum](https://www.twelvelabs.io/legal/dpa): GDPR-aligned DPA for enterprise customers. - [Supplemental Terms](https://www.twelvelabs.io/legal/supplemental-terms): Supplemental terms for specific product surfaces. - [SaaS Vendor Agreement Rider](https://www.twelvelabs.io/legal/saas-vendor-agreement-rider): Rider for enterprise procurement processes. - [AI Standard Clauses](https://www.twelvelabs.io/legal/ai-standard-clauses): Standardized contractual clauses governing AI usage. - [Trademark Usage Guidelines](https://www.twelvelabs.io/legal/tm-usage-guidelines): Brand and trademark usage rules. --- ## Contact - Website: https://www.twelvelabs.io - Sales and partnerships: https://www.twelvelabs.io/contact - Developer Hub: https://www.twelvelabs.io/developer-hub - API documentation: https://docs.twelvelabs.io - Playground (auth required): https://auth.twelvelabs.io/u/login - Trust Center: https://trust.twelvelabs.io - Security contact: security@twelvelabs.io - Careers: https://www.twelvelabs.io/careers - Headquarters: 55 Green Street, San Francisco, CA 94111 - Social: LinkedIn, YouTube, X