About 10 minutes, and you need an installed, running node first (the store is written by the node’s runtime daemon, not by the CLI). You end up with an image recorded in this node’s image store, verified byte by byte on the way in.

Loaded content is runnable. Once the load succeeds and k3sm image ls shows the reference, a Pod naming it runs on the default runtime, and the node materializes the stored layers into the Pod’s filesystem. imagePullPolicy still decides what a reference means, the way it does on any Kubernetes. IfNotPresent and Never serve the loaded content with no registry traffic, while Always, the apiserver’s default for a :latest tag, goes to a registry regardless of what the node holds. Tag loaded images with something other than latest, or set the policy explicitly. Loading stages content for an airgapped node or a pre-seeded cache, and it ships a workload to a node without a registry.

Two Verbs, Two Formats#

There is no format sniffing. A layout handed to load, or a docker-save tar handed to import, is refused with an error naming the other verb.

# a docker-save tar, from k3sm build or from docker save
k3sm build --tag myapp:v1 --output myapp.tar .
k3sm image load myapp.tar

docker save myapp:v1 -o from-docker.tar
k3sm image load from-docker.tar

# a tarred OCI layout
k3sm build --tag myapp:v1 --output ./layout --format oci .
tar -cf layout.tar -C ./layout .
k3sm image import layout.tar

Then look at what the node recorded:

k3sm image ls
k3sm image df

k3sm image also carries push and prune. Run k3sm image --help for the current flags on each.

What Happens to Your Archive#

The CLI opens the archive and streams it to the node’s runtime daemon, which is the store’s only writer. The daemon re-hashes every byte against the digest it is told to expect and records the reference only after that check passes. Nothing is written to the store by the command itself, and the daemon’s socket is readable only by its own account.

Refusals#

	`k3sm image load` and `import`
two tags on one image	rejected. The store records one reference per loaded image, with both tags named in the error; re-save with a single tag, or load once per tag
two images in one archive	rejected for the same reason: one archive, one image
naming	taken from the archive, so `RepoTags` for docker-save and the ref-name annotation for an OCI layout; `--reference` overrides it, and is required for a layout that carries no annotation
signatures and provenance	none evaluated. This is an operator-only surface, and what it stores is exactly what you handed it
a running daemon	required; `load` works against an installed node, not a bare directory
deadline	`--timeout` defaults to 30 minutes for these two verbs, which stream a whole archive, and 2 minutes for the metadata verbs

Next#

serving a model with MLX is the track that will consume this path.
images is the reference page, including the rest of the OCI roadmap.