Bring your own model means an assistant in Nrth OS sends its tasks to a model endpoint you control, on a server in your building, in a data centre you chose, or at a provider you have your own contract with, while the rest of the workspace (assistants, tasks, skills, knowledge, memory) works as before. It is the «Our own model» data policy, set per assistant. It is for companies that already run a model endpoint of their own, or prefer open models they run themselves.
Who needs it
- Companies that already run an open model (Llama, Mistral, Qwen or similar) on their own hardware and want a workspace on top instead of building one.
- Companies with several assistants that want a mixed setup: accounting on their own model, marketing on a hosted EU model.
If none of those is you, «In the EU» or «European providers only» is simpler. Read what EU-hosted means first.
What it requires
| Requirement | Why |
|---|---|
| An HTTPS endpoint with an OpenAI-compatible API | Nrth OS talks to it the same way it talks to hosted providers |
| A model that handles long context and follows structured instructions | Tasks send instructions, knowledge excerpts, memory lines and a skill; a small model will struggle |
| An API key or token | Kept in an encrypted vault, never in the task |
| Enough throughput for your schedules | A Monday-morning run across several assistants is a burst; size for it |
| Someone who owns the endpoint | When it is down, the task fails and tells you; it won’t fall back to a hosted model |
What changes, and what doesn’t
Changes: where the model runs, which model receives the prompts, what it costs you (tasks sent to your own model endpoint are charged at platform price, not model price), and the capability ceiling, which is set by the model you chose.
Doesn’t change: skills, schedules, the feedback loop, the log. Every run still shows what was done and where. The data-policy label on the assistant reads «Our own model».
Capability trade-offs, honestly
Open models can be good enough for much document work. On Thorough tasks, the gap to the strongest hosted models is real. The practical approach is to start a task on a hosted EU model, get the skill right, then switch the assistant’s data policy to «Our own model» and see whether the result holds. Because skills and memory are text, the move is a setting, not a migration.
What about running all of Nrth OS on our hardware?
Not in the first version. Bring your own model covers the model; the workspace itself is hosted in your home region, the EU or the US. Running the whole workspace on your hardware (Nrth Node) is planned for a later version. Companies that need local infrastructure now can get it set up through Nrth AI.
Next step
If you already run a model, apply for the pilot program and tell us the endpoint type in the form.