The Short Answer
Marengo Embed 3.0 is now an embedding model inside Amazon Bedrock Knowledge Bases, so you can point a managed pipeline at video in S3 and get search without handling vectors at all. Two things decide whether that is the right path: whether your search is only over video, and whether those vectors ever need to sit beside embeddings from another model. If both answers are yes and no, use Bedrock. The upgrade catch is the vector itself: 3.0 returns 512 dimensions where 2.7 returned 1024, so a store configured for 2.7 cannot take 3.0 in place.
What actually shipped
Two different things landed and coverage tends to blur them. Marengo Embed 3.0 has been callable as a model on Bedrock since 29 October 2025, through
StartAsyncInvoke on the bedrock-runtime endpoint. What is new, announced by AWS on 10 September 2026, is that it became an embedding model inside Bedrock Knowledge Bases, the managed service. The first is an API you call and then handle the output yourself. The second connects a video library and search works.AWS names iconik among the first customers using Marengo through Bedrock Managed Knowledge Base. That is worth knowing if you run a DAM: the integration you already have may be closer to this than you expect.
Architecture, in one table
| Bedrock Knowledge Bases | Call the model, keep your own store |
| Who holds the vectors | The managed service | You |
| Fuse with embeddings from other models | No | Yes |
| Storage outside AWS | No | Yes |
| Work to first results | Connect a library | Generate, upsert, search |
| Video stays in your account | Yes, in your S3 | Yes, wherever you put it |
Holding the vectors yourself earns its keep on one axis the managed path does not cross: one queryable surface across models and modalities. Marengo vectors beside text embeddings from a different model, beside image vectors from a third, filtered on metadata you own. That is a composition problem rather than a video-search problem.
On benchmarks, plainly
We have not run our own benchmark against Marengo 3.0, so this page publishes no results table. TwelveLabs and AWS have published their own figures and those are theirs to stand behind, not ours to reprint under our name. When we run 3.0 through the same harness as the models already on our video embedding benchmark, the numbers will appear there with the dataset size, the query count and the date attached.
What we can say without measuring anything: the dimension change is the operational fact that bites. Halving the vector halves what the same corpus costs to store, and it also means re-embedding from 2.7 to 3.0 is a new index rather than an in-place write. Read the dimension off your actual response rather than assuming it. Bedrock exposes 3.0 in-Region in us-east-1, eu-west-1 and ap-northeast-2.
Where to go next
The persistence mechanics, the exact
StartAsyncInvoke call and a worked namespace configuration are in How to Store TwelveLabs Marengo Embeddings in Your Own Vector Store.