Marengo 3.0
AI model · Embedding model · Text → Embeddings by Twelve Labs
Multimodal foundation model for video search and any-to-any retrieval.
Marengo 3.0 is a multimodal foundation model for video search and retrieval. It generates 512-dimensional embeddings from video, audio, and text, enabling businesses to build semantic search and recommendation systems.
Model specs
The full spec card (context, pricing per token, modalities and more) on AntSpace. Sign up free to see the full spec card.
- Languages
- English, Korean
Category: Computer Vision
- Modalities
- Video, Audio, Image, Text → Embedding (Vector)
- Embedding dimensions
- 512
- Sizes and variants
- Marengo 3.0, Marengo 3.5
- Access
- API, Amazon Bedrock, Snowflake AI Data Cloud
- Input Modalities
- Video, Audio, Image, Text, Document
- Output Type
- Vector Embedding
- 512-dimensional multimodal embeddings
- Powers semantic and image-to-video search
- Available via API and Amazon Bedrock
- Native integration with Snowflake AI Data Cloud
Overview
Marengo 3.0 is a multimodal foundation model developed by TwelveLabs for video search and any-to-any retrieval. It is designed for developers and enterprises that need to turn raw video footage into searchable, AI-ready data. The model analyzes video frames, temporal relationships, speech, and sound to generate 512-dimensional embeddings. These embeddings power semantic search, allowing users to locate specific moments using natural language or images. Marengo 3.0 is delivered as a managed API and is available through cloud marketplaces such as Amazon Bedrock. It serves as a core engine for high-scale video workflows in industries like media, advertising, and retail.
Search and Retrieval
- Semantic Search: Marengo 3.0 powers semantic search features that allow users to find specific scenes by describing them in natural language.
- Image-to-Video Search: The model enables image-to-video search to find videos semantically similar to a provided image query.
- Multimodal Retrieval: Marengo 3.0 analyzes frames and their temporal relationships along with speech and sound for any-to-any retrieval tasks.
Embeddings and Analysis
- 512-Dimensional Embeddings: The model generates 512-dimensional embeddings that capture the meaning of video, audio, and text.
- Uncertainty Signal: Marengo 3.0 provides an uncertainty signal on every match to help developers gauge the confidence of retrieval results.
- Recommendation Systems Engine: The model can be used to build recommendation systems that understand video content at a semantic level.
- Anomaly Detection: Embeddings can be used to identify anomalies, such as corrupt videos or black backgrounds, to enhance data quality.
Use cases
- Media companies use Marengo 3.0 to search through decades of archive footage using natural language to find specific scenes.
- Retailers integrate the model into recommendation engines to suggest products based on the visual context of videos customers watch.
- Security teams use the model's embeddings to detect behavioral patterns and anomalies across recorded video feeds.
Specifications
- Modalities
- Video, Audio, Image, Text → Embedding (Vector)
- Embedding dimensions
- 512
- Sizes and variants
- Marengo 3.0, Marengo 3.5
- Access
- API, Amazon Bedrock, Snowflake AI Data Cloud
- Fine-tuning
- Custom fine-tuning capabilities are available for specific industry needs upon request.
- Input Modalities
- Video, Audio, Image, Text, Document
- Output Type
- Vector Embedding
- Processing Speed
- Up to 180x real-time indexing
- Cloud Availability
- Amazon Bedrock, Snowflake AI Data Cloud
How it works
- Create an Index: Users create an index and specify Marengo 3.0 as the engine with options for visual and audio modalities.
- Index Video Content: Videos are uploaded to the index where Marengo 3.0 processes frames, audio, and speech to create embeddings.
- Search or Embed: Developers use the Search or Embed endpoints to perform natural language queries or generate vectors for downstream tasks.
Pricing
- Embed Video (Marengo 3.0)
- $2.50 per hour
- Embed Audio (Marengo 3.0)
- $0.50 per hour
- Embed Image (Marengo 3.0)
- $0.10 per 1,000 requests
- Embed Text (Marengo 3.0)
- $0.07 per 1,000 requests
- Index Search (Marengo 3.0)
- $4 per 1,000 queries
FAQ
What is Marengo 3.0?
Marengo 3.0 is a multimodal foundation model that analyzes video, audio, text, and sound to power search and retrieval tasks.
How can I access Marengo 3.0?
Marengo 3.0 is available as a managed API directly from TwelveLabs and is also accessible through cloud platforms like Amazon Bedrock and Snowflake AI Data Cloud.
What is the embedding size for Marengo 3.0?
Marengo 3.0 generates 512-dimensional embeddings that encapsulate the semantics of the video content.
Can Marengo 3.0 be fine-tuned?
Yes, TwelveLabs provides fine-tuning capabilities to help users achieve better results with only a few examples for specific domains.
Added by AntSpace from public sources (twelvelabs.io).