The Deployments page holds the model weights of the team. Promote a draft to Registered before you deploy it. You can also import a checkpoint you trained elsewhere.
Who can do this: project Owners and Editors import checkpoints and promote versions.
| Area | What it does |
|---|---|
| ① Tabs: Weights, Cloud inference, Edge self-hosted | Weights is available. Cloud inference is Partial and Edge self-hosted is Planned. |
| ② Model version card | The version name, its status (Draft or Registered), the bundle status and when it was trained. |
| ③ Import checkpoint | Opens the form that imports a checkpoint as a new model version. |
Import a checkpoint
Choose Import checkpoint
On the Weights tab, choose Import checkpoint.
The Import external checkpoint form asks for these fields. Files go straight from your browser to object storage.
| Field | What to enter | Example |
|---|---|---|
| Project (required) | The project the version belongs to. | — |
| Version name | A name for the version. | e.g. smolvla-so101-v1 |
| Policy family | The policy the checkpoint was trained with. | Select family… |
| Action dimension | The number of action values the robot takes. | e.g. 6 |
| Joint order (comma-separated) | Required when action dimension is set. Enter it on the next step. | e.g. shoulder_pan, shoulder_lift, elbow_flex |
| Train config name (OpenPI only) | Only for OpenPI checkpoints. | e.g. pi0_fast |
Promote and deploy
A draft version cannot be deployed. The Weights tab shows how many drafts must be promoted to Registered before deploy, with a Promote link. Model cards on the Models page show the same status.
Cloud inference is Partial: it needs registered weights and a published inference image. Edge self-hosted is Planned.
Why can't I deploy this version?
The version is still a draft. Promote it to Registered first. A version without a registered status cannot be used for deployment.