You train a model. It works well. You deploy it.
A month later, you retrain with new data. The new model is different. Is it better? Can you compare? What if you need to roll back?
Model versioning answers these questions. Kanva now supports checkpoints.
The versioning problem
In production, models aren’t static. You retrain them as:
- New data becomes available
- Business conditions change
- Features get added or removed
- Bugs get fixed
Each training run produces a different model. Without versioning, you face problems:
No comparison. How does the new model compare to the old one? You can’t tell if you didn’t save the old one.
No rollback. If the new model performs worse in production, can you go back? Only if you kept the previous version.
No audit trail. Which model made that prediction three months ago? For compliance, you might need to know.
No A/B testing. Want to compare models in production? You need multiple versions available simultaneously.
Checkpoints in Kanva
A checkpoint is a saved snapshot of a trained model. It includes:
- The model itself (weights, parameters)
- Training metrics (accuracy, feature importance)
- Feature configuration (which features, how processed)
- Training context (data split, hyperparameters)
Everything needed to reproduce and use that specific model version.
How to use checkpoints
Creating a checkpoint. After training, click “Save Checkpoint” and give it a name. “v1-initial”, “2026-01-retrain”, “before-feature-change”—whatever helps you identify it later.
Viewing checkpoints. The checkpoints panel shows all saved versions with their metrics. Compare accuracy across versions at a glance.
Setting the default. One checkpoint is always the “default”—the one used for predictions unless you specify otherwise. You can change which checkpoint is default at any time.
Using specific checkpoints. Via the API, you can request predictions from any checkpoint, not just the default. Useful for comparison or gradual rollouts.
What gets saved
Checkpoints are self-contained. Each one includes:
| Component | What’s saved |
|---|---|
| Model | Trained weights and parameters |
| Metrics | All training and validation metrics |
| Features | Feature list, types, and engineering config |
| Context | Train/test split, hyperparameters |
| Metadata | Creation date, name, description |
This means you can understand a checkpoint without remembering how you created it. The context is preserved.
Practical workflows
Checkpoints enable several workflows:
Safe retraining. Create a checkpoint before retraining. If the new model is worse, roll back to the checkpoint.
Gradual rollout. Deploy a new model to 10% of traffic while the old checkpoint handles 90%. Compare real-world performance before full switch.
Seasonal models. Some businesses need different models for different seasons. Save checkpoints for each and switch as needed.
Compliance. Regulators ask which model made a decision. With checkpoints, you can point to the exact version.
Experimentation. Try a new feature configuration. Save as a checkpoint. Compare to the previous version. Keep whichever works better.
Checkpoints and the API
The REST API supports checkpoint-specific predictions:
Omit the checkpoint parameter to use the default. Specify it to use a particular version.
This lets you run multiple model versions simultaneously—useful for comparison, testing, or gradual migration.
Getting started
If you have a trained model you want to preserve:
- Open the project
- Go to the Checkpoints section
- Click “Save Checkpoint”
- Give it a meaningful name
The checkpoint is now saved. You can retrain freely, knowing you can always return to this version.
Model checkpoints are available now. Questions? Reach out at hello@human-driven.ai