See the unseen. Know the unknowable.

People don't want AI to do their work. They want to be freed from the mechanical — so they can do the work that matters. Your video archive contains every insight that mattered. Until now, extracting it has been physically impossible.

Sports & Media
Find every comeback moment across our entire 10-year archive
Security & Compliance
Advertising & Publishing
AI Products & Agents
Found 847 comeback moments across 10 years of championship footage. Results ranked by excitement score and audience reaction.

1등

Video-MME 벤치마크 1위 (30분 이상 영상 기준)

+17%

Pegasus 1.5: Gemini 3.1 Pro 대비 17% 높은 성능.

10x

콘텐츠 검토 및 컴플라이언스 스캔 속도 10배 향상

4시간

4시간 분량의 영상도 API 호출 한 번으로 처리

단 몇 분 만에 결과를 확인하세요.

영상 인텔리전스를 위한 인프라.
대규모 원본 영상을 검색 가능하고 AI 활용 가능한 데이터로 전환합니다.

영상 인텔리전스를 위한 인프라.
대규모 원본 영상을 검색 가능하고 AI 활용 가능한 데이터로 전환합니다.

텍스트나
이미지로
장면 검색

자연어로 질문하거나 이미지를 입력하면,
수천 시간 분량의 영상 속
유사 장면을 빠르게 찾아냅니다.

API 및 SDK

프로덕션에 바로 사용할 수 있는 API.
영상을 업로드하고 검색이나 분석을 실행하면 파이프라인에 바로 연동할 수 있는 구조화된 결과를 받을 수 있습니다.
Python 및 Node.js용 공식 SDK는 완전한 타입 지원, 비동기 처리, 재시도 로직, 스트리밍 기능을 제공합니다.

MCP

Claude Code, Cursor, VS Code와 연동되는 MCP 서버로 IDE를 벗어나지 않고도 영상 인텔리전스를 연동하고 활용할 수 있습니다. 멀티모달 데이터을 끊김없이 활용하는 애플리케이션을 구축하세요.

연동

기존 스택은 그대로 두고 트웰브랩스를 바로 연결하세요. 현재 조직에서 사용 중인 도구를 바꾸지 않고도 영상 인텔리전스를 워크플로우에 자연스럽게 통합할 수 있습니다.

영상에 최적화된 인식과 추론, 연동과 활용

대형 언어 모델(LLM)이 텍스트를 연산 가능한 데이터로 만들었듯, 트웰브랩스는 비디오, 이미지, 오디오를 연산 가능하게 만듭니다. 단순한 검색을 넘어 분석과 실행까지 가능하게 합니다.

대형 언어 모델(LLM)이 텍스트를 연산 가능한 데이터로 만들었듯, 트웰브랩스는 비디오, 이미지, 오디오를 연산 가능하게 만듭니다. 단순한 검색을 넘어 분석과 실행까지 가능하게 합니다.

멀티모달 임베딩 모델

보지 못하는 것은 검색할 수 없습니다. Marengo는 영상을 데이터로 바꿉니다. 시공간 임베딩을 통해 누군가 입력한 메타데이터가 아니라, 영상에 실제로 담긴 내용을 기준으로 모든 순간을 검색 가능하게 만듭니다. 하나의 인덱스, 모든 모달리티, 78.5% 복합 정확도, 47개 언어 지원.

영상 언어 모델.

범용 모델은 몇 개 프레임만 보고 추측합니다. Pegasus는 최대 2시간 길이의 영상 전체 시간 흐름을 따라 연속적으로 추론합니다. 시간에 따라 객체, 인과관계, 서사의 흐름까지 이해합니다. 단순히 전사만 읽는 모델이 아닙니다.

영상 언어 모델.

범용 모델은 몇 개 프레임만 보고 추측합니다. Pegasus는 최대 2시간 길이의 영상 전체 시간 흐름을 따라 연속적으로 추론합니다. 시간에 따라 객체, 인과관계, 서사의 흐름까지 이해합니다. 단순히 전사만 읽는 모델이 아닙니다.

영상 언어 모델.

범용 모델은 몇 개 프레임만 보고 추측합니다. Pegasus는 최대 2시간 길이의 영상 전체 시간 흐름을 따라 연속적으로 추론합니다. 시간에 따라 객체, 인과관계, 서사의 흐름까지 이해합니다. 단순히 전사만 읽는 모델이 아닙니다.

Python
Node.js

1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

17

18

19

20

21

import requests

 

# Step 2: Define the API URL and the specific endpoint

API_URL = "https://api.twelvelabs.io/v1.3"

INDEXES_URL = f"{API_URL}/indexes"

 

# Step 3: Create the necessary headers for authentication

headers = {

"x-api-key": "<YOUR_API_KEY>"

}

 

# Step 4: Prepare the data payload for your API request

INDEX_NAME = "<YOUR_INDEX_NAME>"

data = {

"models": [

{

"model_name": "marengo3.0",

"model_options": ["visual", "audio"]

}

]

}

First result: under 10 minutes. No patience required.

Install the SDK, paste your key, point it at a video. It indexes while you read the docs. Free tier ships with 600 minutes. No credit card, no onboarding call, no sales cycle — just results.

First result: under 10 minutes. No patience required.

Install the SDK, paste your key, point it at a video. It indexes while you read the docs. Free tier ships with 600 minutes. No credit card, no onboarding call, no sales cycle — just results.

Others process video. We comprehend it.

General-purpose models stuff frames into a context window at query time — every question starts from scratch. TwelveLabs builds structured knowledge at ingest, so intelligence compounds with every asset indexed.

General-purpose models stuff frames into a context window at query time — every question starts from scratch. TwelveLabs builds structured knowledge at ingest, so intelligence compounds with every asset indexed.

GEMINI (AND GENERAL MULTIMODAL LLMS)

2-minute video cap, 80s audio cap

Gemini's embedding API caps at 2 minutes per video and 80 seconds per audio clip. Anything longer must be chunked manually — destroying temporal context.

Cross-modal embedding collapses

12 distinct audio genres → 1 video result in cross-modal retrieval. Similarity scores cluster between 0.30–0.41 for everything. Calibration fails: unrelated content scores identically to matches.

30.8% structured output failure rate

On news content, Gemini 3.1 Pro fails to produce valid structured JSON nearly a third of the time. Production pipelines require constant fallback handling and retries.

트웰브랩스

Intelligence built at ingest, not at query time

Marengo and Pegasus comprehend every asset the moment it is indexed — building embeddings, structured metadata, and entity relationships that persist. Every query draws on pre-built knowledge, not live inference.

Knowledge that compounds across every asset

Every new asset deepens the knowledge graph. Entity relationships discovered in one video inform retrieval across the entire archive. The system gets smarter with every hour indexed — without you doing anything.

Model upgrades re-process your entire library

When Marengo or Pegasus improves, every asset you've ever indexed automatically gets smarter — no re-upload, no re-indexing, no engineering effort. Historical content becomes more intelligent over time.