Docker Model Runner Blurs Registry, Runtime

Docker Model Runner is close to the right mental model for container-native AI ops, but it stops short of being a true “hosted model repository” for teams that want full control over what is stored, promoted, and retained.

• –Docker’s docs say Model Runner can pull and push models from Docker Hub, OCI registries, and Hugging Face, and package GGUF/Safetensors as OCI artifacts, which makes it registry-friendly.
• –The inference side still depends on a local runner process, so it does not fully collapse into the same artifact-only workflow people expect from a pure registry.
• –Hugging Face models are versioned like git repositories, so you can pin by commit hash or tag for immutability instead of chasing mutable `main` updates.
• –The real gap for this use case is governance: teams want a curated internal store with promotion rules, retention guarantees, and offline availability, not just a cache.
• –For AI infra folks, this lands in the middle of self-hosted model management, not model release hype, which makes it more of an infrastructure question than a launch story.