SageMaker AI now supports SOCI-indexed training and processing images, with AWS benchmarks showing faster startup and shorter job completion times.
Amazon Web Services has added Seekable OCI (SOCI) indexing for Amazon SageMaker AI training and processing jobs, allowing supported containers to start before their full image has been downloaded. AWS announced the capability on October 27, 2026. The service checks for an index when a job starts; if it finds one, it fetches the files needed to launch the container and streams remaining files as they are accessed.
In AWS’s tests, eight compressed container images ranged from 1.8 GB to 35.4 GB. A 13.2 GB image took a median of about six seconds to become ready with lazy loading, compared with three minutes and 42 seconds using a standard download. Across the tested image sizes, AWS reported download-time improvements ranging from eightfold to 46-fold. For the 13.2 GB image, it measured roughly 34-fold faster startup across 13 instance type and size combinations.
The startup improvement does not mean every part of a job runs faster. In AWS’s ResNet-50 image-classification and Qwen2.5-0.5B fine-tuning tests, the first epoch took an additional three to 16 seconds as files were fetched. Later epoch times were unchanged within run-to-run variation, and the tested jobs finished one to six minutes sooner overall. Results may differ when an application reads most of its image before training begins; AWS says that pattern can reduce the benefit.
Teams can use pre-indexed AWS Deep Learning Containers or create an index for a custom image. AWS lists pre-indexed variants for PyTorch Training, TensorFlow Training, vLLM and SGLang, with some arm64 variants. Using the feature does not require changes to the SageMaker job definition, entry point or data channels, and the original image digest and signatures remain valid.
There are limits: images must be in private Amazon Elastic Container Registry repositories, and images under roughly 4 GB may see smaller gains. If an index is unavailable or lazy loading fails, SageMaker falls back to a standard download. AWS says SOCI indexing for SageMaker AI jobs is available in all AWS Regions.