Partnerships

From Footage to Rows: TwelveLabs Pegasus 1.5 Comes to Snowflake

TwelveLabs’ Pegasus 1.5 brings video reasoning to Snowflake AI Data Cloud. Media, advertising, robotics, and industrial teams can find the moments that matter in their footage and put them to use alongside the rest of their data analysis.

Ani Vemprala

Partnerships

From Footage to Rows: TwelveLabs Pegasus 1.5 Comes to Snowflake

TwelveLabs’ Pegasus 1.5 brings video reasoning to Snowflake AI Data Cloud. Media, advertising, robotics, and industrial teams can find the moments that matter in their footage and put them to use alongside the rest of their data analysis.

Ani Vemprala

Nov 18, 2026

6 minutes

Copy link to article

Which moments in a TV show feature a given product? Which scenes best suit a specific advertiser? At what point did a process go out of sequence? The answers to these and similar questions lie inside the petabytes of video a company already stores, yet there hasn’t been a reliable way to find them without a lot of manual effort – or, once found, to bring them alongside the rest of the company’s data.

Today, Pegasus 1.5, TwelveLabs’ video reasoning model, becomes generally available in Snowflake AI Data Cloud. It joins Marengo, our multimodal embedding model, which arrived in June. Together, they let teams work with video where their campaign, operational, and business data already live.

Storing Video is Not the Same as Using It

Video is one of the richest and least readable kinds of data that an enterprise holds. A single two-minute clip carries what is on screen, what is said, what is heard, and how a moment unfolds over time, all at once. That density is part of what makes it so valuable, but a video file has no fields to query, so for a long time, video has mostly been viewed as a storage cost rather than a strategic asset.

Companies have spent years trying to describe their footage after the fact, but most attempts still cover only part of the problem. Manual tagging is slow and varies from person to person. Transcripts capture what was said and miss what happened. Frame-by-frame analysis can spot an object, but loses track of the scene around it.

Teams end up with a thin layer of metadata that still largely describes the file, rather than the content. Ask which scenes were shot at a particular location, or how long a task takes across a month of line recordings, and someone still has to watch hours of footage to find the answer. Closing that gap requires models purpose-built for video understanding.

Two Models, One Workflow

TwelveLabs’ models understand video the way people do. There are two of them, each built for a different job.

Marengo encodes. It reads what is seen, heard, and said, and turns each moment into an embedding that captures its meaning. Similar moments land close together, which is what lets you search footage by describing a scene in natural-language rather than relying on manual tags.

Pegasus reasons. Where most models describe a clip, Pegasus follows it through time: it connects a line of dialogue to a gesture three minutes later, tracks events as they unfold, and interprets why a moment matters.

Here is how that might play out for a streaming team with a catalog of episodes. Marengo finds the scenes that fit a certain description, such as a tense conversation at night. Pegasus then goes through those scenes and describes each one, including the setting, who is on screen, and what happens. Those results land in Snowflake as rows, next to the viewership and ad data already there, allowing the team to answer a question like “do viewers keep watching when the scene is dimly lit?”

Used together inside Snowflake, TwelveLabs’ models turn raw footage into structured intelligence your teams can query like any other data.

What Pegasus 1.5 Brings to Snowflake Cortex AI Functions

Pegasus 1.5 turns hours of footage into data your team can query, moment by moment. Four capabilities make that possible.

Time-based metadata: find every moment that matters

With Pegasus 1.5, you define what matters to your business, whether that’s an editorial segment, a play, a brand appearance, or a speaker change. The model finds every instance, marks where each one begins and ends, and returns timestamped results your downstream systems can read.

No standard benchmark existed for this kind of output, so our team built one. When we ran the tests, Pegasus 1.5 outperforms Gemini 3.1 Pro by 9% on segmentation quality, with boundaries accurate to within roughly 350 milliseconds. That precision is what makes a timestamped result usable in production, since a segment that starts a few seconds late can’t be trusted to cut, score, or flag.

Multimodal prompting: search by showing, not describing

Using a reference image, you can search either for the entity itself (such as a brand’s product) or a person doing something specific (a host interviewing a guest), and Pegasus 1.5 returns every matching moment in the video, timestamped. On our multimodal prompting evaluation, Pegasus 1.5 outperforms Gemini 3.1 Pro by 13.1%.

Two-hour video support: no splitting, stitching, or chunking

Pegasus 1.5 analyzes videos up to two hours long in a single request, keeping context and continuity across the full duration. For teams building on Snowflake, that means not having to deal with chunking logic, and getting one coherent set of results instead of reassembling them from pieces.

Structured output: JSON you can query

Pegasus 1.5 returns JSON that follows the schema you provide, with start and end times, a category, and each requested field, and it holds its structure even when the schema is complex. That reliability is what lets video output join the rest of your business data in Snowflake, with no separate cleanup step.

From Footage to Rows

Once a moment in a video is a row in Snowflake, the analysts who work with your campaign, inventory, and operations data can ask questions of your organization’s video data, in the same place and with the same skills.

Media and advertising teams might want to know things like:

  • Which shot types, speaker changes, or brand appearances line up with drops in viewer retention?

  • Which moments in an episode earn the highest ad completion rates for a given advertiser?

  • How much on-screen time did each sponsor’s logo actually earn, and how does that compare with what they paid for?

  • Which segment types in a creator’s videos bring viewers back for the next one?

Some of the most valuable work in media happens between partners who can’t share raw data with each other, which is what Snowflake’s Data Clean Rooms are built for. Because Pegasus 1.5’s output lands in standard Snowflake tables, content signals like which scenes feature a sponsor’s product can be brought into those workflows alongside audience and outcome data, with no footage shared.

For robotics and industrial teams, segmentation points toward a different set of possibilities. Teams building and running robots record hours of demonstrations, teleoperated sessions, and test runs, and only a few moments in each recording show what the team needs to learn from. A model that can mark where each action begins and ends, and describe what happened in between, lets those teams combine what the video shows with the data they already have in Snowflake, such as sensor logs, task definitions, and outcome records.

They might ask:

  • Which recordings contain the action our robot struggles with most, and how many examples of it do we have?

  • How long does each step of a task take across hundreds of runs, and which test conditions explain the slowest ones?

  • When a sensor logs an error, what was the robot doing on camera at that moment?

  • Which steps most often come right before a failed run or a rework?

These are multiple potential applications for TwelveLabs in Snowflake, but the best questions will always come from the teams who know their own operations the best. The foundation is the same in both fields: accurate boundaries and structured descriptions turn raw footage into something measurable.

Why Bringing Models to the Data Matters

Our approach has always been consistent: bring video intelligence to the platforms where customers already work. When video understanding is built into the platform where the data already lives, it becomes part of the analytics workflow, instead of requiring customers to build, secure, and maintain a separate pipeline. Governance, access controls, and the data that gives video its business context are already in place.

"Video is some of the most valuable unstructured data our customers hold, and some of the hardest to use," said Michelene Rabbit, Global Head of Sports & Music at Snowflake. "For sports leagues, game footage data holds the keys to growing fan engagement and maximizing player health and safety. For media companies, video program data drives distribution and monetization efficiency. By making TwelveLabs’ models available in Snowflake, we’re helping companies across verticals turn video into data they can query, govern, and build on, without moving it somewhere else."

Getting Started

Pegasus 1.5 is generally available in Snowflake AI Data Cloud now. To learn how to set it up with your own video, including how Marengo and Pegasus work together, read the quickstart guide.

Which moments in a TV show feature a given product? Which scenes best suit a specific advertiser? At what point did a process go out of sequence? The answers to these and similar questions lie inside the petabytes of video a company already stores, yet there hasn’t been a reliable way to find them without a lot of manual effort – or, once found, to bring them alongside the rest of the company’s data.

Today, Pegasus 1.5, TwelveLabs’ video reasoning model, becomes generally available in Snowflake AI Data Cloud. It joins Marengo, our multimodal embedding model, which arrived in June. Together, they let teams work with video where their campaign, operational, and business data already live.

Storing Video is Not the Same as Using It

Video is one of the richest and least readable kinds of data that an enterprise holds. A single two-minute clip carries what is on screen, what is said, what is heard, and how a moment unfolds over time, all at once. That density is part of what makes it so valuable, but a video file has no fields to query, so for a long time, video has mostly been viewed as a storage cost rather than a strategic asset.

Companies have spent years trying to describe their footage after the fact, but most attempts still cover only part of the problem. Manual tagging is slow and varies from person to person. Transcripts capture what was said and miss what happened. Frame-by-frame analysis can spot an object, but loses track of the scene around it.

Teams end up with a thin layer of metadata that still largely describes the file, rather than the content. Ask which scenes were shot at a particular location, or how long a task takes across a month of line recordings, and someone still has to watch hours of footage to find the answer. Closing that gap requires models purpose-built for video understanding.

Two Models, One Workflow

TwelveLabs’ models understand video the way people do. There are two of them, each built for a different job.

Marengo encodes. It reads what is seen, heard, and said, and turns each moment into an embedding that captures its meaning. Similar moments land close together, which is what lets you search footage by describing a scene in natural-language rather than relying on manual tags.

Pegasus reasons. Where most models describe a clip, Pegasus follows it through time: it connects a line of dialogue to a gesture three minutes later, tracks events as they unfold, and interprets why a moment matters.

Here is how that might play out for a streaming team with a catalog of episodes. Marengo finds the scenes that fit a certain description, such as a tense conversation at night. Pegasus then goes through those scenes and describes each one, including the setting, who is on screen, and what happens. Those results land in Snowflake as rows, next to the viewership and ad data already there, allowing the team to answer a question like “do viewers keep watching when the scene is dimly lit?”

Used together inside Snowflake, TwelveLabs’ models turn raw footage into structured intelligence your teams can query like any other data.

What Pegasus 1.5 Brings to Snowflake Cortex AI Functions

Pegasus 1.5 turns hours of footage into data your team can query, moment by moment. Four capabilities make that possible.

Time-based metadata: find every moment that matters

With Pegasus 1.5, you define what matters to your business, whether that’s an editorial segment, a play, a brand appearance, or a speaker change. The model finds every instance, marks where each one begins and ends, and returns timestamped results your downstream systems can read.

No standard benchmark existed for this kind of output, so our team built one. When we ran the tests, Pegasus 1.5 outperforms Gemini 3.1 Pro by 9% on segmentation quality, with boundaries accurate to within roughly 350 milliseconds. That precision is what makes a timestamped result usable in production, since a segment that starts a few seconds late can’t be trusted to cut, score, or flag.

Multimodal prompting: search by showing, not describing

Using a reference image, you can search either for the entity itself (such as a brand’s product) or a person doing something specific (a host interviewing a guest), and Pegasus 1.5 returns every matching moment in the video, timestamped. On our multimodal prompting evaluation, Pegasus 1.5 outperforms Gemini 3.1 Pro by 13.1%.

Two-hour video support: no splitting, stitching, or chunking

Pegasus 1.5 analyzes videos up to two hours long in a single request, keeping context and continuity across the full duration. For teams building on Snowflake, that means not having to deal with chunking logic, and getting one coherent set of results instead of reassembling them from pieces.

Structured output: JSON you can query

Pegasus 1.5 returns JSON that follows the schema you provide, with start and end times, a category, and each requested field, and it holds its structure even when the schema is complex. That reliability is what lets video output join the rest of your business data in Snowflake, with no separate cleanup step.

From Footage to Rows

Once a moment in a video is a row in Snowflake, the analysts who work with your campaign, inventory, and operations data can ask questions of your organization’s video data, in the same place and with the same skills.

Media and advertising teams might want to know things like:

  • Which shot types, speaker changes, or brand appearances line up with drops in viewer retention?

  • Which moments in an episode earn the highest ad completion rates for a given advertiser?

  • How much on-screen time did each sponsor’s logo actually earn, and how does that compare with what they paid for?

  • Which segment types in a creator’s videos bring viewers back for the next one?

Some of the most valuable work in media happens between partners who can’t share raw data with each other, which is what Snowflake’s Data Clean Rooms are built for. Because Pegasus 1.5’s output lands in standard Snowflake tables, content signals like which scenes feature a sponsor’s product can be brought into those workflows alongside audience and outcome data, with no footage shared.

For robotics and industrial teams, segmentation points toward a different set of possibilities. Teams building and running robots record hours of demonstrations, teleoperated sessions, and test runs, and only a few moments in each recording show what the team needs to learn from. A model that can mark where each action begins and ends, and describe what happened in between, lets those teams combine what the video shows with the data they already have in Snowflake, such as sensor logs, task definitions, and outcome records.

They might ask:

  • Which recordings contain the action our robot struggles with most, and how many examples of it do we have?

  • How long does each step of a task take across hundreds of runs, and which test conditions explain the slowest ones?

  • When a sensor logs an error, what was the robot doing on camera at that moment?

  • Which steps most often come right before a failed run or a rework?

These are multiple potential applications for TwelveLabs in Snowflake, but the best questions will always come from the teams who know their own operations the best. The foundation is the same in both fields: accurate boundaries and structured descriptions turn raw footage into something measurable.

Why Bringing Models to the Data Matters

Our approach has always been consistent: bring video intelligence to the platforms where customers already work. When video understanding is built into the platform where the data already lives, it becomes part of the analytics workflow, instead of requiring customers to build, secure, and maintain a separate pipeline. Governance, access controls, and the data that gives video its business context are already in place.

"Video is some of the most valuable unstructured data our customers hold, and some of the hardest to use," said Michelene Rabbit, Global Head of Sports & Music at Snowflake. "For sports leagues, game footage data holds the keys to growing fan engagement and maximizing player health and safety. For media companies, video program data drives distribution and monetization efficiency. By making TwelveLabs’ models available in Snowflake, we’re helping companies across verticals turn video into data they can query, govern, and build on, without moving it somewhere else."

Getting Started

Pegasus 1.5 is generally available in Snowflake AI Data Cloud now. To learn how to set it up with your own video, including how Marengo and Pegasus work together, read the quickstart guide.

In this article

No headings found on page

Search, analyze, and explore your videos with AI.

Join our newsletter

Receive the latest advancements, tutorials, and industry insights in video understanding