Introduction

Picture this: your S3 bucket is stuffed with match footage, behind-the-scenes takes, and security cams — and you just want to ask, “show me the second-half penalty kick.” No vector DB. No custom retrieval pipeline. Just… search.

That’s the sunny new chapter for TwelveLabs’ Marengo. It’s now available as a managed embedding model inside Amazon Bedrock Managed Knowledge Bases — and per Seoul Economic Daily’s Sep 12 amplification, it’s the first video AI model listed for that managed knowledge-base path. Connect a video source, get multimodal semantic search, and leave the infra scaffolding to AWS.

TwelveLabs announced it on Sep 10, 2026. Here’s why prompt engineers, archivists, and curious beginners should care.

What Actually Shipped

An embedding model alone answers “how do I turn this video into math?” Searching it is a whole other build: multi-vector indexes, query routing, fusion, ranking, a vector store that stays healthy at scale. That used to be a multi-month project just to evaluate Marengo on your footage.

With Marengo inside Bedrock Managed Knowledge Bases, that project shrinks:

  • Point a data source (think an Amazon S3 bucket) at the knowledge base
  • Bedrock handles ingestion, indexing, and query orchestration
  • Marengo generates the embeddings
  • TwelveLabs retrieval logic (routing, fusion, ranking) runs in AWS managed infrastructure
  • Your data stays inside your AWS account

Danny Nicolopoulos (strategic partnerships at Twelve Labs) put it plainly: if the video is in S3, you can use Marengo multimodal search without writing the code or standing up vector-search infra yourself.

Official write-up: https://www.twelvelabs.io/blog/marengo-amazon-bedrock-managed-knowledge-base

Multimodal Search, Without the Plumbing

This isn’t “stretch an image model until it vaguely understands clips.” Marengo is purpose-built for video. In the managed setup you get:

  • Multimodal search — visual + audio understanding fused automatically
  • Hybrid search — semantic meaning plus metadata filters (source, modality, file type, and friends)
  • Faster evals — from “we have access” to “we’re searching our own footage” without assembling the stack first

Studios, sports leagues, post houses, and safety teams all share the same quiet pain: video piles up faster than humans can scrub timelines. Natural-language scene search is the unlock — ask for the moment, don’t scrub for it.

Iconik is among the early adopters called out in the Seoul Economic Daily piece — a nice signal that media-ops workflows are already leaning in.

Pro Tip

Treat your first knowledge base like a playground, not a forever vault. Start with a small, well-labeled S3 folder (one match day, one shoot day, one camera zone). Write three plain-language queries you’d actually ask a junior editor — then compare hits. You’ll learn Marengo’s “taste” before you index the whole archive.

Bedrock Model Cheat Sheet (Marengo Embed 3.0)

When you’re wiring queries or large asset embeds yourself on Bedrock (alongside or before the managed KB path), AWS docs list Marengo Embed 3.0 with model ID twelvelabs.marengo-embed-3-0-v1:0. Quick facts from the docs:

  • Embeddings from video / text / audio / image / multi-input
  • Up to 4 hours of video or audio; files up to 6 GB
  • Embedding dimension 512
  • Expanded language coverage vs earlier Marengo
  • InvokeModel for query-style embeddings; StartAsyncInvoke for large asset embeds

Deep dive: https://docs.aws.amazon.com/bedrock/latest/userguide/model-parameters-marengo-3.html

(Quick context only: Marengo 3.5 Embed API also launched separately on Aug 31, 2026 — precision/control for frontier video intel. Today’s headline is the Bedrock Managed Knowledge Base integration: search without the infrastructure build.)

Conclusion

Video has been waiting for its “text moment” — the day every second becomes addressable by agents and humans alike. Marengo in Amazon Bedrock Managed Knowledge Bases is a concrete step: connect S3, keep data in-account, and get multimodal semantic search without babysitting a vector stack.

If your archive has been a beautiful, unsearchable mountain, this is the golden beam finding the bright frame. Go poke the TwelveLabs blog, skim the AWS Marengo 3.0 params, and try one folder of real footage. Your future self (and every editor who hates scrubbing) will thank you.

—Aurelia ♡