Model Artifact

A model artifact is a saved representation of a model that can be loaded or used by another process. Depending on the framework, it may contain weights, a serialized model, or a package of files. A usable release also needs compatible code, preprocessing, configuration, and runtime dependencies, whether packaged with the artifact or referenced separately.

The weights are not always the whole model

Imagine saving a text classifier but omitting the tokenizer vocabulary that assigned numbers to words. Loading the weights with a different vocabulary can produce incorrect results even when tensor sizes match. Record which components belong together.

MLflow's model format illustrates packaging model files with metadata and environment information. It is one convention rather than a universal file format.

Identity and usability are separate checks

An immutable version or content digest can identify the exact saved bytes. It does not prove that the model meets quality requirements or runs correctly in a new environment. Check expected inputs and outputs, dependency compatibility, and representative loading behavior.

A registry tracks versions and associated records. Deployment places an accepted artifact and its dependencies into an execution environment. Serving uses that environment to provide results. Keeping these roles distinct helps explain why uploading a model file is not the same as operating a prediction service.

References

• MLflow: Model packaging