TwelveLabs released Pegasus 1.6 on October 6, 2026, a video intelligence model designed for egocentric video analysis. The model is built to support footage captured from a first-person perspective, such as wearable cameras or teleoperated systems, which is essential for training robotics and physical AI.
The model automates labor-intensive tasks like action segmentation, labeling, and dense captioning of spatial interactions. It also includes quality control features to flag issues in footage. According to the company, manual labeling of such video can take 70 to 155 hours per hour of footage, a process the new model aims to scale.
Pegasus 1.6 features a 261,120-token context window, allowing for the processing of videos up to 2 hours long. It also supports the analysis of 1 to 20 images per request and provides improved entity recognition. Access to the model is provided through the TwelveLabs API.
While the new version builds on the foundation of Pegasus 1.5, which launched on April 20, 2026, some limitations remain. High-volume bulk processing and batch analysis are currently still handled by the previous version, Pegasus 1.5.
