Pegasus 1.5

TwelveLabs Pegasus 1.5 Overview

TwelveLabs Pegasus 1.5 is a proprietary video language model designed to convert long-form video into structured, queryable understanding. Newly available through Amazon Bedrock, it is notable for focusing less on generic chat and more on deep video comprehension: identifying what happens in a video, when it happens, and how those events can be represented as reliable metadata. Its headline advance is the combination of temporal localization, structured output, and multimodal prompting, making it useful for teams that need machine-readable video intelligence rather than simple captions or summaries.

Pegasus 1.5 is built for video-understanding workflows such as content indexing, media search, compliance review, highlight generation, sports and entertainment analysis, e-learning navigation, and surveillance-style event review where permitted by policy and law. The model can segment videos into meaningful scenes or events, attach timestamps, describe visual and spoken content, and return outputs in JSON for downstream systems. It also supports prompts that include reference images, enabling use cases such as finding scenes visually similar to a product, logo, person, object type, interface state, or visual style.

Key Capabilities

Core capabilities include multimodal video analysis, temporal grounding, metadata generation, segmentation, and structured-output generation. Instead of treating a video as a single blob, Pegasus 1.5 can reason across time and produce timestamped descriptions, chapters, labels, or event boundaries. Its JSON output support is particularly important for production use: developers can request predictable schemas for clips, entities, actions, topics, confidence-like fields, and time ranges, making the model easier to integrate with search indexes, asset-management systems, review queues, and analytics pipelines.

The model’s reference-image prompting expands its usefulness beyond text-only queries. A user can supply an image as a visual target and ask the model to locate related moments in a video, describe matching segments, or compare visual context. This makes Pegasus 1.5 especially relevant for large media libraries where visual similarity and scene-level retrieval are more important than transcript search alone.

Technical Specifications and Availability

Pegasus 1.5 supports video and image inputs with text instructions, and it generates text or structured JSON outputs. It is available through Amazon Bedrock, allowing customers to access it through Bedrock’s managed API, security controls, and cloud deployment infrastructure. TwelveLabs has not publicly positioned Pegasus 1.5 as an open-weight model; it should be treated as proprietary/commercial. Pricing, quota limits, regional availability, maximum video duration, context window, token limits, and maximum output size should be verified in the current Amazon Bedrock and TwelveLabs documentation, as these details can vary by region, account configuration, and service tier.

Strengths and Benefits

Pegasus 1.5’s main strength is specialization. General-purpose multimodal models can describe frames or short clips, but Pegasus is optimized for long-form video understanding and timestamped retrieval. That makes it more suitable for operational workloads involving large archives, media review, or structured video intelligence. Bedrock availability is also a practical benefit for enterprises already using AWS, since it can simplify access control, monitoring, and integration with existing cloud systems.

Limitations and Caveats

As with any video language model, outputs should be validated for high-stakes uses. The model may miss brief events, infer intent incorrectly, struggle with low-quality footage, or produce imperfect timestamps when scenes are ambiguous. Structured JSON can improve integration, but it does not guarantee factual correctness. Costs may also become significant for long videos or large archives, and proprietary access means users cannot self-host or inspect weights.

Compared with broad multimodal systems such as Gemini, GPT-class vision models, or Claude-style document-and-image assistants, Pegasus 1.5 is more narrowly focused on video intelligence and temporal metadata. Compared with earlier TwelveLabs offerings, the emphasis on Bedrock availability, reference-image prompting, segmentation, and structured JSON makes it better suited for enterprise-scale video understanding pipelines.