What it is
Marengo 3.5 is the multimodal embedding model from TwelveLabs, a company that specialises in video understanding, launched on 31 August 2026. It turns video, audio, images, text and visual PDF pages into vectors in one shared space, so a text or image query can find the matching moment inside hours of footage. Marengo 3.0 remains the model behind TwelveLabs' hosted /search endpoint, which does not support 3.5.
Vectors default to 512 dimensions and can be cut to 256 or 128; TwelveLabs says 256 keeps 98.3% of the full score and 128 keeps 94.6%. Queries can mix text with an image, video or audio clip, timestamped text such as scripts or player names can be fused into the media embedding, and an optional uncertainty signal helps an application decide when to rerank or abstain. Video has no duration cap, against four hours for 3.0, and the docs list 36 languages.
Billing has moved from minutes of video to tokens charged per modality: at launch $0.26 per million video tokens, $0.065 audio, $0.08 image, $0.162 document and $0.20 text. Account plans run Free, Plus ($29 a month), Pro ($99) and Enterprise, each with storage and a monthly workload allowance; all accounts move to this plan structure on 4 January 2027.
Our review
Marengo 3.5 belongs on the shortlist whenever the corpus is mostly video, but the case rests on TwelveLabs' own tests. It reports a mean of 70.57 across five text-to-video retrieval datasets, 6.29 points above Marengo 3.0, and 67.38 on an eight-group aggregate at 512 dimensions against 62.35 for Gemini Embedding 2 and 50.34 for Amazon Nova Multimodal Embeddings at 3,072. Marengo has no entry on the MTEB video tasks, so nobody independent has checked.
Vector size is the practical advantage. By TwelveLabs' account a 128-dimension vector still scores 63.75 on that aggregate, above both rivals at 24 times the size, and vector databases bill roughly in proportion to dimensions. An archive stored at 128 rather than 3,072 dimensions needs about a twenty-fourth of the space.
Cost is harder to predict than it should be. The per-token rates are published, but the pages we read do not say how many tokens a second of video or audio produces, and plan allowances are described only as Basic, Medium and High. A video with sound is billed for both tracks. Budgeting a 10,000-hour archive therefore means embedding a sample and reading the usage report.
Two traps sit in the docs. Marengo 3.5 vectors cannot be mixed with 3.0 vectors, so upgrading means re-embedding everything, and the hosted /search endpoint still runs on 3.0, so 3.5 users bring their own vector database. On data, TwelveLabs' privacy policy says customer data is used to improve its services with no opt-out, deferring to customer contracts where it acts as processor, and its security and compliance pages name no certification. Qwen3-VL-Embedding, which can be self-hosted, sidesteps both questions at the cost of video-specific tuning.
Where it earns its place
A search engineer at a sports broadcaster with a 20-year match archive
Let producers find a specific play by describing it
Embed matches asynchronously with timestamped scores and player names fused in, store 128- or 256-dimension vectors in their own database and query with text. Anything embedded on 3.0 must be redone, and the bill depends on a token count per hour of video that TwelveLabs does not publish.
A media-asset manager at a 30-person ad agency
Find reusable shots across past campaigns from a reference frame
Composed queries combine an uploaded image with text such as 'same angle, at night'. Pro at $99 a month includes 200 GB of storage and optional overage, which suits a library of a few hundred hours; Plus at $29 stops all usage at its limits.
A trust-and-safety lead at a video-sharing platform
Flag uploads that closely resemble known prohibited clips
Embed new uploads, compare them with a reference set and send borderline matches to a person using the uncertainty signal. TwelveLabs says the signal is not a probability of correctness, so thresholds need tuning on labelled traffic before it drives removals.
Before you commit
- Setup
- Create a TwelveLabs account, call the synchronous Embed endpoint for queries and the asynchronous task API for video, audio, images and PDFs. The output is vectors you store yourself, since /search does not support 3.5.
- Integrations
- TwelveLabs REST API and SDKs. Inputs: video in any FFmpeg format from 360x360 to 5184x2160, JPEG or PNG images up to 32 MB, WAV, MP3 or FLAC audio, PDFs up to 512 MB. AWS Bedrock availability for 3.5 is not stated on the pages we read.
- Data and privacy
- TwelveLabs' privacy policy (January 2026) says customer data and outputs, including embeddings, are used to improve its services, with no opt-out; customer contracts govern where it acts as processor. Retention lasts as long as the account is used. No certification is named on its security or compliance pages; Enterprise adds SSO/SAML.
- Cost at scale
- At launch rates, one billion video tokens cost $260 and the matching audio $65, but the token count per hour of footage is not published, so archive costs need a test run. Plans add storage: 50 GB on Plus at $29 and 200 GB on Pro at $99.
- Limits
- Video must be at least 4 seconds, with no upper limit; images at least 128x128. PDFs are embedded page by page, other document types are 'coming soon'. Rate limits depend on upload method and are not on the model page. 3.5 vectors are incompatible with 3.0.
What's good
- Built to find moments in video, not just frames
- 512-dimension vectors, down to 128 with about 5% loss
- Composed queries and timestamped metadata fusion
- No video duration cap; 36 languages
What's not
- All benchmarks are TwelveLabs' own
- Tokens per second of video not published
- Not supported by TwelveLabs' own /search endpoint
- Privacy policy lets customer data improve its services
Best for
Teams building search over large video libraries who want a managed, video-specialist embedding API and will run their own vector store.
Alternatives
Marengo 3.5: common questions
- Is Marengo 3.5 worth it?
- Purpose-built for finding moments in video, with small vectors that keep storage cheap, though every number behind it comes from TwelveLabs. We rate it 3.9 out of 5.
- How much does Marengo 3.5 cost?
- Embed API at launch, per 1M tokens: video $0.26, audio $0.065, image $0.08, document $0.162, text $0.20. Plans: Free, Plus $29/month, Pro $99/month, Enterprise custom.
- What are the downsides of Marengo 3.5?
- All benchmarks are TwelveLabs' own; tokens per second of video not published; not supported by TwelveLabs' own /search endpoint; privacy policy lets customer data improve its services.
- Who is Marengo 3.5 best for?
- Teams building search over large video libraries who want a managed, video-specialist embedding API and will run their own vector store.
- How much does Marengo 3.5 cost for a team or in production?
- At launch rates, one billion video tokens cost $260 and the matching audio $65, but the token count per hour of footage is not published, so archive costs need a test run. Plans add storage: 50 GB on Plus at $29 and 200 GB on Pro at $99.
- Is Marengo 3.5 safe for company data?
- TwelveLabs' privacy policy (January 2026) says customer data and outputs, including embeddings, are used to improve its services, with no opt-out; customer contracts govern where it acts as processor. Retention lasts as long as the account is used. No certification is named on its security or compliance pages; Enterprise adds SSO/SAML.
- What are the best Marengo 3.5 alternatives?
- The closest alternatives we have reviewed are Gemini Embedding 2, Qwen3-VL-Embedding and Nemotron ColEmbed V2.
Answers checked Oct 22, 2026.