Public alpha · Apache-2.0
Open models.
One runtime.
One open-source CLI that finds, verifies, and runs the right engine for each model.
Validated: Windows x64 Linux x64 preview macOS arm64 build preview

$ backpack pull smollm2-135m$ backpack run smollm2-135mready · llama.cpp · local
One interface, not another engine
Bring the model.
Backpack brings the rest.
Give Backpack a packaged model. It selects the pinned adapter, verifies the model and runtime artifacts, and owns the session.
Not another inference engine.llama.cpp, whisper.cpp, and isolated Python workers stay underneath one interface.
No global engine setup.Runtime bundles are pinned, installed into Backpack data, and verified before launch.
Experimental agent launch
Keep the tool.
Change the model.
Start an installed coding agent against Backpack's local API. Provider settings stay isolated to that child process.
$ backpack launch claude$ backpack launch codexBring the agent yourself. Launch compatibility is experimental; no catalog model is external-agent-qualified yet.
$ backpack pull smollm2-135m$ backpack run smollm2-135m --detach$ backpack ps$ backpack stop <session-id>$ backpack transcribe sample.wav --model whisper-large-v3-turbo$ backpack speak "Hello" --model kokoro-82m --output hello.wav$ backpack launch codexLocal execution is the validated path. User-owned SSH compute remains experimental ↗.
Execution-validated
Models that
actually run.
Supported means pull, verify, launch, infer, and stop are exercised end to end.
Full compatibility matrix ↗Experimental code/vision packages and package-change-required image/video contracts are intentionally not presented as supported.
Early alpha
Pull it.
Run it.
The installer selects the newest published release, warns for prereleases, verifies the archive checksum, and installs without elevation.
Inspect-first and pinned installs ↗
Then
$ backpack doctor
$ backpack models
$ backpack pull smollm2-135m
$ backpack run smollm2-135m