# Models and inference

Separate model lifecycle, inference, and training.

Source: https://docs.minds.sh/docs/concepts/models



A model is a component used to transform inputs into predictions, embeddings, or generated outputs. Minds exposes several interfaces around models, so begin by choosing the operation you need.

| Operation               | Purpose                                            |
| ----------------------- | -------------------------------------------------- |
| Model lifecycle         | Inspect, download, or manage a model artifact      |
| Embedding               | Turn supported input into a numeric representation |
| Generation or inference | Run a configured model on an input                 |
| Training                | Update learned parameters for a task               |

## A model name is not enough [#a-model-name-is-not-enough]

The backend, model artifact, hardware, configuration, and input schema must agree. A model listed in the workspace does not prove it is loaded, ready, or compatible with every inference route.

For embeddings, record the model and dimension used to build an index. For inference, retain the request settings needed to reproduce the result. For training, keep evaluation and recovery separate from a simple job acknowledgment.

Use [model management](/docs/api/instance/model-management), [machine learning](/docs/api/instance/machine-learning), and [unified inference](/docs/api/instance/inference) to find the specific registered contracts. Do not assume a route is OpenAI-compatible unless the reference explicitly establishes that behavior.