The Repository (Models module → Repository) is the catalogue layer: the place to discover models your organisation can deploy and to bring new ones into the workspace.
What the Repository shows
- Available base models — models your platform can serve locally, usually open-weight model families across sizes and capability tiers.
- Downloaded/imported models — what already exists in your environment.
- Bring-your-own entries — custom weights your organisation has added.
Use it to answer: do we already have something suitable, or do we need to add one?
Bringing in a model
Depending on your deployment's policy, new models arrive in one of these ways:
| Route | When it's used |
|---|---|
| Pull from the catalogue | Request a listed base model and the platform fetches and serves it. |
| Import a configuration | Bring in a custom model definition (for example from another environment's export). |
| Bring your own weights | Provide model files your organisation owns (BYOM). Usually performed with administrator help. |
Note:
Model downloads consume disk space, bandwidth and GPU memory. In restricted environments, model acquisition is administrator-controlled — see Models (Admin) and Cluster Manager → Storage for capacity.
From base model to custom model
A base model is generic. To make it yours:
- Find the base model in the Repository.
- Create a custom model from it (rather than editing the base directly).
- In the editor, set the system prompt, parameters, tags, prompt suggestions, vision toggle, documents and tools. See Models Overview.
- Test it in chat; compare it against the base with Model Comparison.
- If it still isn't good enough for your domain, fine-tune it — see Fine-Tuning.
Choosing well
| Requirement | What to look for |
|---|---|
| Fast, high-volume simple tasks | Small model, low latency |
| Complex reasoning, long documents | Larger model with a long context window |
| Reading screenshots, diagrams, scans | A model with vision support |
| Strictly factual, cited answers | Any strong model plus a knowledge base |
| Specialised terminology/format | Consider fine-tuning |
Guidance: start with the smallest model that meets quality when tested on your real examples. Bigger costs more compute for everyone on the shared cluster.