Skip to main content
Open Models in your project and add a model. The registry records what was trained and how it should be loaded; it does not download or execute the artifact.

​
Model details

Provide the base model’s repository ID and revision, artifact type and location, training method, and the data used. An adapter needs its matching base model; a checkpoint contains the saved model weights. Add a checksum where available. Record the tokenizer, chat template, system prompt, and generation settings used in evaluation. Differences here can change the response even when the weights match. Use artifact references without passwords, access tokens, or signed URL query strings. Credentials belong in the runtime configuration, not the registry.

​
Draft and registered versions

You can edit a draft. Registration creates an immutable configuration with a fingerprint. Incomplete or unverified details remain visible; registration alone does not prove that files exist or are compatible.

​
Connect a runtime

R3AL currently configures serving runtimes for pilot projects. The runtime must load the expected model and report the matching identity and fingerprint. Only then can the project show Ready. A model may be registered while the project still says Not connected. That means the serving connection is missing, not that another registration is needed.

​
Model support

The current hosted insurance demo is a configured Qwen checkpoint. Registering another model name does not establish that its architecture, adapter, quantization format, memory requirements, or license are supported. Confirm compatibility before planning a deployment. Self-service training, pruning, quantization, and LoRA/QLoRA execution are being integrated separately. No training job starts when you register a model.