Models
The trainable artifacts — custom architectures, foundation models, and your own uploads, all in one catalog.
A model is a trainable artifact: an architecture plus configuration
that a pipeline can train on your data. predictAI
keeps every model in one catalog; three kinds, distinguished by a single
kind field:
kind | What it is | Where it's documented |
|---|---|---|
custom | An architecture you define (TensorFlow, PyTorch, scikit-learn, XGBoost) | Custom models |
foundation | A pretrained foundation model — predictfm, Chronos-Bolt, TimesFM, … | Foundation models |
byom | A pre-trained model you upload | BYOM |
Thirty seconds of API
# Browse the public catalog (all kinds, one list)
curl "$API_BASE/v1/models/public?kind=all&limit=5" \
-H "Authorization: Bearer $TOKEN" \
-H "X-Workspace-Id: $WORKSPACE_ID"
# Deploy a foundation model zero-shot — no training required
curl -X POST "$API_BASE/v1/models/deploy-foundation" \
-H "Authorization: Bearer $TOKEN" \
-H "X-Workspace-Id: $WORKSPACE_ID" \
-H "Content-Type: application/json" \
-d '{ "family": "predictfm", "slug": "predictfm-f-v0.1" }'Internally, model definitions are sometimes called blueprints — you may see that word in response payloads (e.g. an ID field). In the API paths and all prose, they're always just models.
In this section
Concepts
Model kinds, models vs pipelines, and catalog visibility.
Custom models
Create your own architecture and browse the public catalog.
Foundation models
Browse pretrained models, fine-tune adapters, or deploy zero-shot.
BYOM
Upload, validate, and register a model you trained elsewhere.
A model on its own doesn't train or serve anything. Training it on your data is a pipeline; putting a result live is a deployment.