What is a video embedding model?
A video embedding model turns a stretch of video into a vector, a list of numbers placed so that similar footage lands close together. Search compares vectors: a query is embedded the same way and the nearest segments come back. If the model also has a text encoder trained into the same space, a sentence can find footage, and if it accepts images, a single frame can. The model decides what a search can see: some read only pixels, some read frames over time, and some read the audio track as well.
This page lists the models teams use in 2026 with their limits and prices, shows a text query and a frame query against one set of video embeddings that you can run with no key, and covers what goes wrong.
Which video embedding models are there in 2026?
Facts below come from each vendor's own documentation and pricing pages on October 18, 2026. "Not stated" means the vendor does not publish it.
| Model | Input per embedding | Dimensions | Access and price | Notes |
|---|---|---|---|---|
| TwelveLabs Marengo 3.0 | Up to 4 hours per file, fixed segments of 1 to 10 s (default 6) | 512 | Hosted. $2.50 per hour of video from TwelveLabs; on Amazon Bedrock $0.0007 per second ($2.52 per hour) | Separate visual, audio and transcription vectors plus an optional fused one |
| TwelveLabs Marengo 3.5 | Video, audio, image, text and documents; segments dynamic or fixed | 128, 256 or 512 | Hosted. $0.260 per million tokens for video; tokens per second not stated | Vectors are not compatible with 3.0, and the search endpoint does not support 3.5 yet |
| Google Vertex multimodalembedding@001 | Video segments with a minimum interval of 4 s; the default covers the first 120 s in 16 s intervals | 1408 for video | Hosted. Per embedding: $0.002 for 4 to 8 s intervals, $0.001 for 8 to 15 s, $0.0005 for 15 s and longer | The model behind the example on this page |
| Google Gemini Embedding 2 | One video per request, up to 80 s with audio or 120 s without | 128 to 3072 (default 3072) | Hosted. $12.00 per million video tokens online, $6.00 in batch; tokens per second not stated | One space for text, images, video, audio and documents |
| Amazon Nova Multimodal Embeddings | Up to 30 s of video plus 30 s of audio; the async API segments longer files | 256, 384, 1024 or 3072 | Bedrock. $0.0007 per second of video ($0.042 per minute) | Audio and video can share one vector or keep two |
| Voyage multimodal-3.5 | Video files up to 20 MB; duration and sampling not stated | 256 to 2048 (default 1024) | Hosted. $0.60 per billion pixels, each frame billed as an image | Video limits are thinly documented |
| Qwen3-VL-Embedding (2B and 8B) | 1 frame per second, up to 64 frames | Up to 4096 (8B) | Open weights, Apache-2.0 | The 8B model reports 57.0 on MMEB-V2 video retrieval; serving an 8B model costs more than a dual encoder |
| InternVideo2 | 4 to 8 frames | Not stated | Open code under Apache-2.0; weights gated on Hugging Face | Reports 50.0 to 55.9 MSR-VTT text-to-video recall@1, zero-shot |
| VideoPrism | 8 or 16 frames at 288 px | Not stated | Open weights, Apache-2.0 | The LvT-B model reports 50.1 MSR-VTT text-to-video, zero-shot |
| LanguageBind Video | 8 frames | Not stated | Open weights, MIT | Video, audio, depth and text in one space |
| X-CLIP | 8 frames at 224 px | 512 | Open weights, MIT | A light 2022 model that still works for text-to-video |
| SigLIP 2, per frame | One image per frame | 1152 | Open weights, Apache-2.0 | Strong and cheap on what is visible; no motion or audio |
| Meta V-JEPA 2 | 64 frames | 1408 | Open weights, Apache-2.0 | No text encoder, so no text-to-video search on its own; suited to classification |
| Jina v5 omni | Frames, optionally with audio | 768 or 1024 | Weights under CC BY-NC 4.0, so no commercial use | Jina reports video as its weak spot (27.8 on MMEB video retrieval) |
OpenAI has no video embedding model: text-embedding-3-small, text-embedding-3-large and ada-002 accept text only.
The benchmark numbers in the table are each vendor's own claims, measured under different protocols (frame counts, zero-shot or fine-tuned, re-ranking), so they do not rank the models against each other. Test candidates on your own footage and queries.
How do I choose one?
Five questions settle most choices.
1. Will people search with words? Then the model needs a text encoder in the same space. V-JEPA 2 has none; every other row above does. 2. How long is each piece you embed? Nova takes 30 seconds per embedding, Gemini Embedding 2 takes 80 to 120, and Vertex prices by interval length. Segment length also sets how precisely a result can point at a moment. 3. Does sound matter? Marengo, Nova, Gemini Embedding 2 and LanguageBind read audio. A per-frame image model sees none of it. 4. Hosted or self-run? Hosted models cost per minute and keep no GPUs busy. Open weights cost nothing per call and need GPUs and someone to run them; check the license, since Jina v5 omni rules out commercial use. 5. What does an hour of footage cost? At list price: Marengo 3.0 is $2.50 an hour, Nova is $2.52, and Vertex depends on the interval. Cut into 6-second segments, Vertex is 600 embeddings an hour at $0.002, or $1.20; at 15-second intervals it is 240 at $0.0005, or $0.12.
Worked example: a sentence and a frame against the same video embeddings
The public footage library behind the footage intelligence template holds 67 clips (63 from Pexels, 4 from NASA) cut into 6-second segments, each embedded with Vertex multimodalembedding@001 at 1408 dimensions. Two public retrievers search that one set of vectors: one takes text and one takes an image. Neither needs an API key.
A text query:
curl -s -X POST https://api.mixpeek.com/v1/public/retrievers/footage-intelligence-sample/execute \
-H 'Content-Type: application/json' \
-d '{"inputs":{"query":"someone tying their shoelaces"},"pagination":{"method":"offset","page_number":1,"page_size":3}}'A frame as the query, taken 14 seconds into the shoelace clip:
curl -s -X POST https://api.mixpeek.com/v1/public/retrievers/footage-intelligence-frame-search/execute \
-H 'Content-Type: application/json' \
-d '{"inputs":{"image_url":"https://storage.googleapis.com/mixpeek-gtm-assets/footage-intelligence/frames/pexels-3804690-t14.jpg"},"pagination":{"method":"offset","page_number":1,"page_size":3}}'What each returned on October 18, 2026:
text: "someone tying their shoelaces" 7.0s to 13.0s score 0.1405 Person tying shoelaces 0.0s to 7.0s score 0.0978 A man on the starting line of a running track 0.0s to 7.0s score 0.0409 A man pouring ground coffee on a tin cup filter image: frame at 14 s of the shoelace clip 13.0s to 19.0s score 0.9145 Person tying shoelaces 7.0s to 13.0s score 0.6014 A man on the starting line of a running track 0.0s to 7.0s score 0.5303 Close up on make up pallette and brushes
Both queries find the right clip first, and both land on the same runner-up, because a runner at a starting line is the nearest footage to shoelaces in a small library. The frame lands in the 13 to 19 second segment, which holds the second it was taken from.
The scores show the trap. A correct text match scores 0.14 and a correct frame match scores 0.91, against the same vectors. Text-to-video and image-to-video similarities live on different scales, and they also shift between models. A threshold has to be tuned for each pair of modalities and each model, and it has to be tuned again when the model changes.
Model upgrades mean re-embedding. Marengo 3.5 vectors cannot be searched against 3.0 vectors, and Marengo 2.7 is being retired. Every model change is a re-index of the whole library, so keep the source files in your own storage. See changing embedding models without re-embedding everything at once.
Per-frame models miss motion and sound. SigLIP or CLIP on sampled frames sees what is visible in each frame and nothing about what happens between them. Add a transcript index for speech, or use a model that reads audio.
Segment length sets precision. A 30-second embedding can only point at a 30-second window. Short segments point precisely and cost more embeddings per hour.
Sampling misses short events. At one frame a second, a 0.3-second event may never be seen. See how frame sampling changes what video embeddings see.
One vector per video is too coarse. A single embedding for a 20-minute video blends every scene together. Embed segments and group the results by video at query time.
It depends on what you search with and where you run it. For a hosted API with sound and long files, Marengo 3.0 and Amazon Nova cost about the same per hour. For cheap visual search on Google Cloud, Vertex multimodalembedding@001 at longer intervals is the lowest-priced row above. For open weights, Qwen3-VL-Embedding-8B reports the strongest video retrieval of the open models here, at a higher serving cost than InternVideo2 or VideoPrism.
No. OpenAI's embedding models accept text only. Teams on OpenAI usually caption or transcribe each segment and embed that text, which loses everything the caption leaves out.
Yes, frame by frame: sample frames, embed each with CLIP or SigLIP, and store the frame's video and timestamp beside the vector. It works well for what is visible and misses motion, sound and speech.
At October 2026 list prices: Marengo 3.0 $2.50, Amazon Nova $2.52, and Vertex multimodalembedding@001 from $0.12 (15-second intervals) to $1.20 (6-second intervals). Open models cost the GPU time to run them. Storage and search of the vectors are separate costs, covered in how much a vector database costs.
Most models fix it or offer a few sizes: 512 for Marengo 3.0, 1408 for Vertex, up to 3072 for Nova and Gemini Embedding 2. Smaller vectors are cheaper to store and search and usually lose a little accuracy; test the smaller size on your own queries before committing to it.
Sample library and its index settings: the footage intelligence sample, built October 15, 2026, and its frame search retriever, tested October 18, 2026.
TwelveLabs Marengo 3.5, migrating from 3.0 to 3.5 and overage prices
Marengo 3.0 on Amazon Bedrock and Amazon Bedrock pricing
Vertex AI multimodal embeddings and Vertex AI pricing
Gemini Embedding 2
Amazon Nova multimodal embeddings
OpenAI embeddings
Voyage multimodal embeddings and Voyage pricing
Jina embeddings v5 omni
Model cards: Qwen3-VL-Embedding-8B, V-JEPA 2, VideoPrism, X-CLIP, LanguageBind Video, SigLIP 2 and the InternVideo2 model zoo
The scores show the trap. A correct text match scores 0.14 and a correct frame match scores 0.91, against the same vectors. Text-to-video and image-to-video similarities live on different scales, and they also shift between models. A threshold has to be tuned for each pair of modalities and each model, and it has to be tuned again when the model changes.
What goes wrong with video embeddings?
Frequently asked questions
What is the best video embedding model?
It depends on what you search with and where you run it. For a hosted API with sound and long files, Marengo 3.0 and Amazon Nova cost about the same per hour. For cheap visual search on Google Cloud, Vertex multimodalembedding@001 at longer intervals is the lowest-priced row above. For open weights, Qwen3-VL-Embedding-8B reports the strongest video retrieval of the open models here, at a higher serving cost than InternVideo2 or VideoPrism.
Does OpenAI have a video embedding model?
No. OpenAI's embedding models accept text only. Teams on OpenAI usually caption or transcribe each segment and embed that text, which loses everything the caption leaves out.
Can I use CLIP to embed video?
Yes, frame by frame: sample frames, embed each with CLIP or SigLIP, and store the frame's video and timestamp beside the vector. It works well for what is visible and misses motion, sound and speech.
How much does it cost to embed an hour of video?
At October 2026 list prices: Marengo 3.0 $2.50, Amazon Nova $2.52, and Vertex multimodalembedding@001 from $0.12 (15-second intervals) to $1.20 (6-second intervals). Open models cost the GPU time to run them. Storage and search of the vectors are separate costs, covered in how much a vector database costs.
How many dimensions should a video embedding have?
Most models fix it or offer a few sizes: 512 for Marengo 3.0, 1408 for Vertex, up to 3072 for Nova and Gemini Embedding 2. Smaller vectors are cheaper to store and search and usually lose a little accuracy; test the smaller size on your own queries before committing to it.