GPU-accelerated PyTorch text and video embedding service for NVIDIA Cosmos Dataset Search. Model weights are provisioned separately.

Cosmos Embed1 OSS
Cosmos Embed1 OSS is the GPU-accelerated text and video embedding service used by NVIDIA Cosmos Dataset Search. Release 1.2.0 provides a CVDS-owned PyTorch inference service for the Cosmos Embed1 model.
The service exposes embedding APIs for text and video, batch requests, health checks, and operational statistics. Video processing uses GPU decoding through PyNvVideoCodec and eight-frame sampling for the 224p model.
Deployment
Run the service on a supported NVIDIA GPU system with the NVIDIA Container Toolkit. Model weights are provisioned separately and are not included in this container. The deployment supports a mounted model snapshot and an offline inference configuration.
For deployment instructions and API documentation, see the Cosmos Dataset Search repository.
Third-party software and separately obtained model weights retain their applicable licenses.