Twelve Labs Interview Guide (2026): Video AI Rounds and What to Actually Prep

Twelve Labs is a San Francisco company, with a large engineering and research presence in Seoul, building foundation models that understand video. It was founded in 2021 by Jae Lee (CEO), Aiden Lee (CTO), and Soyoung Lee along with early co-founders, and it ships two model families through an API: Marengo, a video embedding model for search and retrieval, and Pegasus, a video-language model for summarization, classification, and question answering over footage. In July 2026 the company raised a $100 million Series B co-led by NEA and NAVER Ventures, with Amazon, NVIDIA’s NVentures, Index Ventures, Radical Ventures, and others backing it across roughly $200 million total to date. Customers use the API to search, tag, and describe large video libraries, which shapes what its engineering and research teams hire for.

Process

Public detail on the Twelve Labs loop is thin, so treat this as the likely shape based on stage and domain rather than a fixed script. Expect a recruiter screen first (background, why video AI, why a fast startup), then a technical phone screen, then a virtual onsite of three to five rounds that usually splits by track. Research and ML roles go deep on multimodal and video modeling plus an ML system design discussion, with a coding round to confirm you can implement and not just describe. Backend and infrastructure roles put more weight on systems, serving, and data pipelines, still with a strong coding bar. Both tracks tend to close with a hiring-manager or founder conversation that carries the behavioral signal about ownership, speed, and how you work through hard trade-offs. Cycle length runs from roughly two to five weeks from first call to decision, sometimes faster when the team is set on a candidate they want.

What they actually ask

• Video-language modeling: how a model like Pegasus turns raw footage into text (captions, summaries, answers), and where it breaks down on long or low-signal video.
• Multimodal embeddings: how Marengo-style joint embeddings align video, audio, speech, and text so a plain text query retrieves the right clip.
• Video retrieval at scale: design a search system over millions of hours of embeddings, covering approximate nearest-neighbor indexes, sharding, and tail latency.
• Temporal modeling: how you represent time in video through frame sampling, temporal attention, and localizing a specific moment inside a long clip.
• Large-scale video data pipelines: ingesting, decoding, and preprocessing large volumes of video for training and inference without your compute budget exploding.
• Inference for video plus a coding bar: serving heavy models cost-effectively (batching, GPU utilization, caching embeddings), plus medium-level coding to confirm you write clean, correct code under time.

Levels and comp (2026)

Twelve Labs is well funded and pays like a competitive SF AI startup, but public comp data is thin, so treat any single figure with caution. Total compensation is base plus equity, sometimes with a bonus, and equity typically vests over four years with a one-year cliff. For a Series B company at this stage, expect base salaries in a broad band that rises with level from mid to senior to staff, plus a meaningful equity grant whose real value depends on the outcome rather than the headline number. Research and ML roles at the top of the market can push higher on both base and equity than a typical backend role, and Seoul-based positions are usually banded to that local market instead of the Bay Area one. Because private-company bands move fast after a fresh round, check levels.fyi and recent Blind threads for current numbers, and ask the recruiter directly for base, equity units, and the latest preferred-share price so you can value the offer yourself. Do not over-index on paper equity value; ask about strike price, vesting, and the last round’s terms before you sign.

Prep priorities

Match your prep to the track. If you are aiming at research or ML, be ready to explain video and multimodal architectures end to end: how frames get sampled and encoded, how video and text share an embedding space, how temporal information survives, and where current video models still fall short. Read the Twelve Labs blog and the Marengo and Pegasus write-ups so you can talk about their approach and not a generic vision model. If you are aiming at backend or infra, drill system design for video (embedding stores, nearest-neighbor search, streaming ingestion, GPU serving) and keep your coding sharp at medium level. Both tracks reward people who reason out loud, so narrate your trade-offs as you go instead of jumping to an answer. Prepare two or three project stories with real decisions you owned, and have a clear reason why video understanding and a fast startup fit you right now.

Frequently Asked Questions

Is the Twelve Labs interview more ML research or engineering? It depends on the role. Research and ML positions weight multimodal and video modeling, while backend and infrastructure positions weight systems, serving, and data pipelines. Both share a coding bar, so confirm your track with the recruiter and prep the matching depth.

How hard is the interview? Public difficulty data is thin, but expect a bar in line with a well-funded AI startup: medium coding done cleanly, plus real depth in either video and multimodal ML or large-scale systems. The domain knowledge is usually what separates candidates, more than the raw coding.

Do I need video or multimodal experience to get hired? It helps a lot for research and ML roles, and it is a plus for infra roles that serve these models. You can offset gaps with strong general ML or systems skills and clear evidence that you have studied how video models actually work.

How long does the process take? Plan for roughly two to five weeks from first contact to offer, sometimes faster when the team is set on a candidate. Timelines vary with the role, the location (San Francisco or Seoul), and scheduling.