Concepts
Model kinds, how models relate to pipelines, and how catalog visibility works.
Model kinds
Every model document carries a kind discriminator. The three kinds
share the same catalog fields (name, description, visibility,
categories, tags, rating, in_use_by, …) and differ in the
kind-specific body:
kind | Kind-specific body | How it trains |
|---|---|---|
custom | framework, model_blueprint, compilation_details, training_details, data_split, data_preprocessing | Full training from scratch on your segment |
foundation | foundation: family, slug, task, intensity, lookback, horizon | Fine-tunes an adapter on a foundation model |
byom | byom: framework, model_uri, input_signature, output_signature | Doesn't train — you already trained it |
Two things to know about foundation models:
- Every foundation model is identified by
family+slug(predictfm/predictfm-f-v0.1,chronos_bolt/chronos-bolt-base). Its pretrained weights are frozen — they never change. - An adapter is a small set of weights fine-tuned on your data and
applied on top of a foundation model. Training a
foundation-kind model produces an adapter; the foundation model itself never changes.
Models vs pipelines
The distinction the whole API leans on: a model is what trains; a pipeline is how and when it trains. A model has no data, no schedule, and can't serve predictions until a pipeline binds it to a segment and runs a training. One model can back many pipelines.
Read the full comparison (with a diagram) in Models & pipelines.
Model routes use an internal name in their paths: they say blueprint
(/v1/models, /v1/models). The paths
are verbatim; the prose vocabulary here is what the platform means.
Public catalog & visibility
Every model is either private (visible to its owner and workspace) or
public (listed in the shared catalog for every workspace). You choose
at creation time and it's part of the model document:
{ "visibility": "public" }The public catalog —
GET /v1/models/public
— is one merged list across all kinds. Foundation models appear
in it as catalog rows too, with synthesized ids:
| Row | uid shape | kind in the catalog row |
|---|---|---|
| Foundation model (predictfm) | pfm__predictfm-f-v0.1 | foundation_backbone |
| Foundation model (other families) | extfm__chronos_bolt__chronos-bolt-base | foundation_backbone |
Foundation-model rows are always public. Fetching one of these ids from
GET /v1/models/{uid} returns the same catalog shape as a
regular model, plus a foundation object with model-specific fields
(weights_uri, eval_metrics, license, …).
Foundation models and adapters have their own three-level visibility
(private / org / public) because they can be shared inside an
organization as well as publicly — see
Foundation models.
Quotas
Each plan caps models per workspace; creating past the cap returns 403.
BYOM models count against a separate byom_models quota, checked at
register time. Quotas and what
each one gates live in Billing.