News

Snowflake adds generally available TwelveLabs video AI to Cortex functions

Snowflake customers can use TwelveLabs' Marengo 3.0 and Pegasus 1.5 through Cortex AI Functions to embed, retrieve and analyze video with SQL; the reviewed sources do not establish broad adoption.

D
Nov 18, 2026 · 2 min read

Snowflake has added TwelveLabs’ Marengo 3.0 and Pegasus 1.5 to Cortex AI Functions. Customers can now use SQL functions to embed, search and analyze video alongside other data in Snowflake. The company described the combined workflow on November 18.

Each model handles a different stage. Marengo 3.0 works through AI_MULTI_EMBED, converting staged video and text queries into 512-dimensional vectors that represent visual, audio and transcribed-speech content with timestamps. Teams can compare those vectors to retrieve relevant segments, then pass a selected file to Pegasus 1.5 through AI_COMPLETE for summarization, question answering, event identification or structured extraction. Snowflake’s multimodal guide documents that search-then-analysis workflow.

The November announcement brings together capabilities that reached general availability in two steps. Snowflake made Marengo-based semantic video search generally available on August 10 and added generally available Pegasus 1.5 support on November 16. General availability describes product status; the reviewed sources provide no independent customer-count, usage-volume, production-scale or revenue data showing broad adoption.

Snowflake documents AI_MULTI_EMBED support for video or audio files of up to four hours and 6 GB from internal or external stages, with transcribed-speech and text-query support across 36 languages. Its Pegasus release note sets a two-hour and 2 GB limit for video analysis and lists MP4, MOV, AVI, MKV and WebM as supported formats.

Snowflake says video-derived data kept in its tables can use row-level policies, role-based access controls and audit logs. It also says the SQL workflow avoids a separate video platform and, depending on the use case, may remove the need for standalone frame-extraction or transcription pipelines, while TwelveLabs describes Marengo’s embeddings as designed to reduce processing and storage demands. The reviewed sources include no independent benchmark or workflow-specific security audit for those claims.

Snowflake cites Challenger School’s marketing team as a customer example. The company says the team built a footage-search tool with AI_MULTI_EMBED that returns timestamped results and transcript context in minutes instead of a manual process lasting weeks. The outcome appears in Snowflake’s announcement and was not independently corroborated in the reviewed evidence.

More news