TwelveLabs
by Twelve Labs Inc.
Video understanding models and APIs that make video libraries searchable and queryable by content.
TwelveLabs sells video foundation models through an API: Marengo produces multimodal embeddings for search and retrieval, and Pegasus generates text and structured metadata from video. It is aimed at media companies, sports leagues, advertisers and public-sector teams that hold large video archives and currently rely on manual tagging or thin metadata to find anything in them.
TwelveLabs is a San Francisco company, with research in Seoul, that builds foundation models for video and exposes them through a REST API with official Python and Node.js SDKs. The platform has two model families. Marengo is a multimodal embedding model: Marengo 3.0, released on November 30, 2025, produces 512-dimensional embeddings across video, audio, image and text, supports composed queries that combine an image with text, and handles non-Latin languages such as Korean and Japanese. Those embeddings power the Search API (natural-language search across an indexed library) and the Embed API (vectors you store in your own database). Pegasus is a video-language model: Pegasus 1.5, generally available on April 20, 2026, takes a user-defined JSON schema and returns timestamped, structured results across a video of up to two hours in a single call, with no indexing step, and accepts a reference image to find every appearance of a given entity. The vendor reports a 13.1% segmentation-quality gain for Pegasus 1.5 over Gemini 3.1 Pro on multimodal prompting. Both model families are also offered through Amazon Bedrock, and Enterprise customers can get private or self-managed deployments. Named customers on the vendor site include NFL Media, MLSE, Sejong City, Voxel51 and MindsDB. In July 2026 the company raised a $100 million Series B co-led by NEA and NAVER Ventures, with Amazon participating. It competes with general multimodal models such as Google Gemini and Amazon Nova, and with older label-based video APIs, by specialising in long-form video and pricing per hour of footage.
An engineering or media-operations lead at a broadcaster, sports league or agency who owns thousands of hours of video and needs to search or tag it by what is actually on screen and in the audio.
Natural-language search and schema-defined, timestamped metadata across a video archive without hand-tagging footage.
At a Glance
- Category
- AI Models & APIs
- Pricing
- Freemium, Usage-based, Contact for pricing
- Target Market
- CTOs, Media & Entertainment Engineering Teams, Enterprise Developers, Data Scientists, Public Sector
- Deployment
- API-based, Cloud-first, Self-hosted
- Headquarters
- San Francisco, United States
Key Features
- ✓Marengo 3.0 embeddings
Produces 512-dimensional multimodal embeddings across video, audio, image and text, which keeps vector storage costs low for large archives.
- ✓Search API
Runs natural-language queries against an indexed video library and returns the matching moments, removing the need to scrub footage manually.
- ✓Pegasus 1.5 time-based metadata
Returns structured, timestamped JSON against a schema you define, for videos up to two hours, without an indexing step.
- ✓Multimodal and composed queries
Accepts a reference image alongside text so users can find every moment a specific person, logo or object appears.
- ✓Embed API
Exports embeddings for video, audio, image, text and documents so teams can run retrieval inside their own vector store.
- ✓Amazon Bedrock availability
Marengo and Pegasus are offered as Bedrock models, letting AWS customers use them under existing cloud contracts and controls.
- ✓Private and self-managed deployment
The Enterprise plan offers private or self-managed deployment for organisations that cannot send footage to a shared SaaS service.
Capabilities
Use Cases
- •Sports archive search
A league or broadcaster indexes years of game footage and lets producers find specific plays or players with plain-language queries.
- •Structured metadata for media asset management
Pegasus fills a custom JSON schema with timestamped scenes, people and events, populating a MAM system without manual logging.
- •Ad and content compliance review
Teams screen long-form video for brand, rights or policy issues and get flagged timestamps for human reviewers to check.
- •Highlight and clip generation
Editors locate the key moments in a long event recording automatically and cut highlight packages faster than manual review allows.
- •Public-sector and security video analysis
Agencies search large volumes of recorded footage for specific events or objects, as in the vendor's Sejong City deployment.
- •Video RAG for agents
Developers store Marengo embeddings in their own database so an agent can retrieve and cite relevant video segments when answering questions.
Ideal For
Best For
- ✓Semantic search across large broadcast, sports or archive video libraries
- ✓Generating timestamped structured metadata (scenes, entities, events) to a custom JSON schema
- ✓Building video retrieval or RAG pipelines on your own vector database using exported embeddings
- ✓Teams already on AWS that want video models available through Amazon Bedrock
- ✓Automated highlight creation, segmentation and compliance review of long-form content
Not Ideal For
- ✗Teams that need to generate or edit video footage; TwelveLabs analyses video and does not create it
- ✗Buyers with only occasional, short clips, where a general multimodal model they already pay for may be cheaper to run
- ✗Organisations that need fixed list pricing for private deployment, since Enterprise terms are custom-quoted
Integrations
Deployment
Market Analysis
Pros
- ✓Published per-hour and per-token pricing with a free tier that needs no credit card
- ✓Covers both retrieval (Marengo) and structured metadata generation (Pegasus) behind one API
- ✓Available through Amazon Bedrock as well as direct SaaS, with a self-managed option for Enterprise
- ✓Named production customers in sports and media such as NFL Media and MLSE
Cons
- ✗Independent review coverage is thin: Product Hunt shows a single review and G2 data could not be retrieved
- ✗A hackathon participant on Hacker News (Oct 2025) reported a hard request limit with no visible counter on free accounts
- ✗A Feb 2025 Hacker News post questioned whether Pegasus 1.2 reasoned over video directly or relied on a transcription and captioning pipeline; the claim is unverified but reflects sceptical practitioner reception
- ✗Benchmark claims against Gemini and Amazon Nova are vendor-published and were not independently replicated in the sources reviewed
Pricing
Free
$0
- ✓Up to 600 minutes of cumulative video indexing
- ✓Search, Analyze and Embed APIs at no charge
- ✓Indexes kept 90 days, 100 videos and 10 hours per index
- ✓5 concurrent tasks, no credit card
Developer
Usage-based (indexing $2.50/hour of video)
- ✓Search API $4 per 1,000 queries
- ✓Analyze API $1.75/hour of input video plus $7.50 per 1M output tokens
- ✓Embed API from $0.065 per 1M audio tokens to $0.260 per 1M video tokens
- ✓Infrastructure $0.09/hour
- ✓Unlimited indexes, up to 25 concurrent tasks
Enterprise
Contact for pricing
- ✓Custom duration, volume and concurrency limits
- ✓Private or self-managed deployment
- ✓Model fine-tuning on request
List pricing is published for the pay-as-you-go Developer plan and is metered per hour of video indexed or analysed, per 1,000 search queries, and per million embedding or output tokens, plus an hourly infrastructure charge. The Free plan caps cumulative indexing at 600 minutes. Private deployment, higher concurrency and fine-tuning sit behind custom Enterprise pricing. Amazon Bedrock usage is billed through AWS instead.
Security & Compliance
Connect
Sources
This page was written from 10 sources, 6 on domains other than twelvelabs.io.
- 1.twelvelabs.io — twelvelabs.iovendor
- 2.twelvelabs.io — pricingvendor
- 3.twelvelabs.io — marengo 3 0vendor
- 4.twelvelabs.io — twelvelabs series b 100mvendor
- 5.finsmes.com — twelvelabs raises 100m in series b funding
- 6.prweb.com — twelvelabs launches pegasus 1 5 turning raw video into struc
- 7.aws.amazon.com — twelvelabs
- 8.github.com — twelvelabs io
- 9.hn.algolia.com — hn.algolia.com
- 10.producthunt.com — twelvelabs
Stay Ahead of the Curve
Weekly enterprise AI insights for technology leaders. No spam, no vendor pitches—unsubscribe anytime.
SubscribeRelated Products
Arena AI Evaluations
Paid model evaluations built on the Arena leaderboard's crowdsourced human votes, for enterprises choosing an LLM
Kimi K3
2.8T-parameter open-weight MoE model with a 1M-token context, priced at $3/$15 per million tokens
TypeSafe Jev
A 'System One' decision model that returns typed choices, scores and calibrated probabilities instead of text, for agent routing and classification
Xiaomi MiMo-V2.6
MIT-licensed, 1M-context multimodal MoE models that took the top open-weights score on Artificial Analysis at well under $1 per million tokens