Model Version

What Is a Model Version?

A model version is the runnable form of a model. It captures everything required to execute the model consistently, including:

Build artifacts

Code and dependencies packaged for this version.

Weights

The trained parameters that define the model’s behaviour.

Runtime environment

Framework, language runtime, and hardware integration.

Entry points

How tasks are executed for this version.

Why Versions Exist

Models evolve. Versions let you:

Release updated behavior without breaking existing integrations
Test experimental changes safely

Immutable by design

Versions cannot be changed after creation. Any modification results in a new version.

How Versions Affect Predictions

When you run a model:

WriftAI routes the request to one of its versions
That version runs the prediction and produces outputs

Version selection may be:

Explicit — you pass a version number
Implicit — WriftAI uses the model’s latest version

Identifying a Version

Each version has a unique version number.

To refer to a specific version of a model, you use the combined form: owner/name:version-number

Example:

john/image-upscaler:30

Here:

john/image-upscaler identifies the model
30 is the unique version number
The combined reference points to one exact runnable version of that model

Schemas

Schemas make a version self-describing — your app always knows what to send and what to expect back. WriftAI's Web Interface also uses them to shape the model playground experience.

Task-Level Schemas

Every task has two schemas:

Input schema

Defines what your request must provide. Field types, required fields, defaults, and allowed values are specified here.

Output schema

Describes what the task returns, including fields, types, and formats of the result.

Examples

For the prediction task, a version exposes:

prediction.input
prediction.output

For versions that support training:

training.input
training.output

Schema Format

Schemas use JSON Schema Draft 2012.