dmr

dmr

dmr is the standalone Docker Model Runner: a single binary that bundles both the inference daemon and the full model management CLI, with no dependency on Docker Desktop or a running Docker Engine.

dmr serve              # start the daemon (foreground)
dmr pull ai/gemma3     # pull a model
dmr run ai/gemma3 "Hi" # run it
dmr ls                 # list local models
dmr ps                 # list running models
dmr rm ai/gemma3       # remove a local model

How it fits together

• dmr serve runs pkg/server directly, in-process — the same daemon used by the docker-model-runner container and by Docker Desktop's bundled runner. It listens on TCP port 12434 by default (or a Unix socket via --socket); see dmr serve --help.
• Every other dmr subcommand is the same command tree as the docker model CLI plugin (cmd/cli/commands), so it has full feature parity (run, pull, push, tag, inspect, logs, bench, configure, ...). See cmd/cli/README.md for full command documentation.
• dmr always talks to its daemon over MODEL_RUNNER_HOST (default http://localhost:12434), which forces the "manual host" code path in cmd/cli/desktop/context.go. This is what makes dmr engine-independent: unlike docker model, it never probes a Docker Engine connection or checks for Docker Desktop.
• Docker-Engine-only commands that manage a docker-model-runner container (install-runner, start-runner, stop-runner, restart-runner, reinstall-runner, uninstall-runner) are hidden from dmr, since dmr serve replaces them — there's no container to manage.

Building

make build-dmr         # native build, output ./dmr (or dmr.exe on Windows)
make build-dmr-cross   # cross-compile every published target into dist/dmr/<os>-<arch>/

dmr has no cgo dependencies, so it cross-compiles cleanly for macOS (arm64), Linux (amd64/arm64), and Windows (amd64) from any host. See ../../packaging/README.md for how these builds are released via Homebrew (brew install docker/tap/dmr) and WinGet (winget install Docker.dmr).