Pegasus 1.5 on Amazon Bedrock can identify requested video segments and return timestamped metadata, including from videos up to two hours long.
AWS published details on October 19, 2026, about TwelveLabs’ Pegasus 1.5 video-language model on Amazon Bedrock. The article notes that availability had been announced on October 9. The model can analyze a video against user-defined criteria and return matching sections as structured, timestamped data.
A key addition is time-based metadata generation. Rather than asking a general question about a clip, users can define the kinds of segments or events they want to find, then receive matching segments with start and end times and requested fields. The article gives examples including editorial sections, sports plays, speaker changes, and brand appearances. Segment definitions can combine text with reference images to help identify a particular person, product, or logo in context.
Pegasus 1.5 can process videos up to two hours in one request, twice the one-hour limit AWS gives for Pegasus 1.2. Videos can be supplied from Amazon S3 without a pre-indexing step, or as an inline base64 string for clips up to 25 MB. The model is listed for on-demand inference in US West (Oregon), with model ID `twelvelabs.pegasus-1-5-v1:0`; requests use the Bedrock InvokeModel API, with a streaming option also available.
The service retains capabilities such as video description, timestamping, and analysis of visual and audio content. AWS lists full English support and partial support for Arabic, Chinese, French, German, Italian, Japanese, Korean, Portuguese, Russian, Spanish, Thai, and Vietnamese. Users can request free-form text or constrain output to a JSON Schema for downstream processing.
For teams handling large video libraries, the practical change is that specific scenes can be located and represented as structured data without first building an indexing pipeline. AWS says source videos can be sent from S3 or inline, and that data remains within the AWS data boundary.