# Amazon Bedrock adds Marengo 3.0 embeddings for media search > AWS says Marengo 3.0 brings video, audio and image embeddings to Bedrock Managed Knowledge Base, with 512-dimensional vectors and segment timestamps. Canonical URL: https://www.devobs.io/news/news-event-2b624b2ffccd4b0c55be08a8/ By: Ines Costa Published: 2026-09-12T22:22:22.801Z Updated: 2026-09-12T22:22:22.801Z Event date: 2026-09-11 Section: AI AWS announced on September 11, 2026 that TwelveLabs Marengo 3.0 is available as an embedding model in Amazon Bedrock Managed Knowledge Base. According to the [announcement](https://aws.amazon.com/about-aws/whats-new/2026/09/amazon-bedrock-managed-knowledge-base-multimodal-embeddings-twelvelabs-marengo/), the integration extends the service’s existing media-search approach of transcribing audio and video and embedding the resulting text. ## What the integration adds AWS says Marengo 3.0 creates multimodal embeddings for video, audio and images, encoding visual scenes, speech and video cues directly rather than relying on transcription alone. The company describes a workflow in which customers upload media from data sources such as Amazon S3, synchronize the knowledge base and search using natural language. The model produces 512-dimensional vectors and offers configurable segmentation. AWS also says results include segment start and end times, enabling applications to take users to a relevant moment in a video. These capabilities are described in the [AWS announcement](https://aws.amazon.com/about-aws/whats-new/2026/09/amazon-bedrock-managed-knowledge-base-multimodal-embeddings-twelvelabs-marengo/). ## What engineers should evaluate The useful evaluation question is whether the integration finds content that transcript-based search misses. Test queries that describe visible actions or scenes—not just words spoken in the footage—and compare the results with an existing search pipeline. That would help distinguish the value of multimodal retrieval from familiar transcript matching without assuming an accuracy improvement. Segmentation deserves its own tests. Check whether the returned intervals contain enough context and whether their timestamps provide useful navigation points. Use representative media rather than treating vector dimensions as a proxy for retrieval quality. The 512-dimensional output is an input to index planning, not a measure of end-to-end accuracy or total operating cost. ## Availability AWS describes the integration as available. Its September 11 announcement does not explicitly label the release as preview or general availability, and it does not specify supported AWS Regions or pricing. Teams should confirm those deployment details before adopting the integration. ## Source references -