Marengo 3.5
Embedding modelCard summarizedMarengo 3.5 is an embedding model. It analyzes video, audio, images, and documents.
Where to call it
- TwelveLabs API
marengo3.5
Good for
- Similarity search, classification, clustering, recommendations and RAG over video, audio, images and documents
- Composed queries, e.g. a player photo plus a clip plus a text description in one vector
- Attaching play-by-play, scripts or rundowns to the moments they describe (time-based metadata fusion)
Overview
TwelveLabs · 02Marengo, front
Marengo 3.5
Marengo 3.5 is an embedding model. It analyzes video, audio, images, and documents.
Typical price
$0.20 / 1M tok
Text input
- Dimensions
- 512
- Max input
- No duration limit for async video/audio
- Modes
- video embeddings, audio embeddings
- Released
- 31 Aug 2026
Marengo 3.5 is TwelveLabs' video-native multimodal embedding model. It puts video, audio, images, text and PDF pages in one 512-dimensional space, with Matryoshka truncation to 256 or 128. It adds composed queries that mix text with video, audio or image references, fusion of your own timestamped text into segment embeddings, an optional per-dimension uncertainty vector, and no duration cap on async video and audio. It is an embedding model only: platform search still runs on Marengo 3.0. Launch posts (31 Aug 2026); Marengo 3.5 docs page; release notes
marengo3.5- Typical price
- $0.20 / 1M tok
- Dimensions
- 512 (default; 256 or 128 via Matryoshka)
- Max input
- No duration limit for async video/audio; sync: 30 s video/audio, 2,000 text tokens, 32 MB per media source; documents ≤ 512 MB
- Languages
- 36 languages, including English
- Weights
- Closed
- Released
- 31 Aug 2026
- video embeddings
- audio embeddings
- image embeddings
- text embeddings
- visual PDF and text document embeddings
- multimodal embeddings (composed text + video/audio/image queries)
- any-to-video retrieval (in your own vector store)
- time-based metadata fusion
- embedding uncertainty