Downloading Model Artifacts
When training completes or a checkpoint is exported, the model artifacts are stored in the platform's object storage. You can download them for offline analysis or to load the model in your own environment.
Prerequisites
| Requirement | Details |
|---|---|
| Role | FAE or R&D engineer |
| Model state | The model version has been registered (training completed or checkpoint exported) |
Archive Contents
Model artifacts are provided as a .tar.gz archive. When extracted, an ACT model, for example, looks like this:
text
pick_place_v1_v20261020-01/
├── model_card.json # Model card: lineage, dataset versions, step, model family
├── checkpoint/
│ └── model.safetensors # Model weights
├── config.json # Model architecture and hyperparameter snapshot
├── normalization.json # Normalization statistics for observations and actions
├── inference.yaml # Inference config: cameras, control frequency, action chunking
└── manifest.sha256 # Checksums for the artifact files| File | Purpose |
|---|---|
model_card.json | Records the training job ID, dataset versions, model family, step, progress (for checkpoints), and export time |
checkpoint/ | Weight files |
config.json | Architecture parameters needed to load the model |
normalization.json | Used at inference time to normalize observations and denormalize actions |
inference.yaml | Inference parameters such as input cameras and control frequency |
manifest.sha256 | Used to verify the download is complete |
Files may differ slightly between model families; model_card.json is the authoritative description.
Downloading
- Go to Model Deployment → Model Management and find the model.
- Click Details and select the version to download.
- Click Download; your browser downloads the artifact archive.
- After the download finishes, verify its integrity:
bash
tar -xzf pick_place_v1_v20261020-01.tar.gz
cd pick_place_v1_v20261020-01
sha256sum -c manifest.sha256The download is complete if every file reports OK.
Reading the Model Card
python
import json
with open("model_card.json", encoding="utf-8") as f:
card = json.load(f)
print(card["model_family"], card["step"], card["dataset_versions"])Usage Notes
- Use the same normalization: Always use the
normalization.jsonfrom the archive at inference time; otherwise the action outputs will be distorted. - Match cameras and frequency: The camera configuration and control frequency at inference time should match
inference.yaml. - Validate on a robot first: We recommend completing on-robot validation through NexCore deployment and evaluation before using the model elsewhere.
- Keep the lineage: When distributing a model, include
model_card.jsonso its training data and version can be traced.
Common Issues
| Issue | Resolution |
|---|---|
| No Download button | Your role doesn't have download permission, or downloads are disabled by platform policy |
| Checksum verification fails | Download again. If it still fails, submit a support request with the model version. |
| Loading fails with an architecture mismatch | Make sure you're using the config.json from the archive and code for the matching model family |

