One thing we’ve seen is that Git LFS starts to become painful once AI projects accumulate hundreds of checkpoints or many dataset versions. Even if you optimize git lfs pull
, you’re still managing a Git repository that wasn’t originally designed as an artifact management system.
For source code, Git + LFS works well. For AI artifacts, many teams eventually move toward a dedicated artifact platform where models, checkpoints and datasets are versioned independently from Git. That avoids a lot of unnecessary transfers and keeps repositories much smaller. We’re actually building CAVS Node around this idea—content-addressable storage with artifact versioning and deduplication for AI workloads. The goal isn’t to replace Git, but to let Git manage code while CAVS manages large immutable artifacts. It has been a much cleaner workflow in our experiments.