Skip to main content
Evidence state (2026-09-17, v0.35.0): Private AI is experimental. The current public release declares its components (Ollama and Open WebUI), and no automated lifecycle run covers install, apply, verify, backup, or restore yet. Runtime evidence is pending.
Private AI runs Ollama and Open WebUI on your Linux Docker host. Ollama serves the model; Open WebUI provides the chat interface. The local owner receives the initial administrator account, public signup is disabled, and no kombify account is required. The use case is experimental. Check the release catalog for its declared components.

​
Check your host capacity first

The CPU profile has a declared host floor of 4 CPU cores, 12 GiB RAM, and 40 GiB disk; 8 CPU cores, 16 GiB RAM, and 80 GiB disk are recommended. Model size matters on top of that floor:
  • Ollama has a declared 6 GiB RAM limit — the downloaded model weights plus the context you use must fit inside it.
  • Open WebUI has a declared 2 GiB RAM limit.
  • Allow disk space for container images and for downloaded model weights together.
  • GPU serving is not enabled; a higher compute tier uses the same CPU runtime.
These limits describe capacity, not response speed. Check the available host resources and the chosen model’s requirements before installation. See Choose a StackKit for compute profiles.

​
Select the AI workload

From an empty owner workspace:
The init command creates the AI configuration and its private credential references. When adding AI to an existing workspace, update the StackSpec and run stackkit secrets materialize. Review and apply the configuration:
Run Apply on the supported Linux host using its observed inventory; see Review, plan, and apply.

​
Open the AI route as the owner

After apply, open the generated AI route. Open WebUI creates its initial administrator from the established local owner’s identity and the owner-password secret that the init command placed in owner custody. From your StackKits workspace, retrieve the initial password in a private terminal:
  • Keep the displayed password private and do not copy it into chats, screenshots, or Git.
  • Change the password inside Open WebUI after the first login. The startup bootstrap does not reset an existing account, so keep your changed administrator password in your own custody.

​
Choose and download a model explicitly

No model is selected or pulled for you — not during install, restart, or upgrade. In Open WebUI, open Admin Settings, explicitly select an Ollama model, and request its download. Before you pick one, review the upstream model’s license, download size, quantization, context length, and RAM requirements. Start with a small model on the CPU profile. Larger models and longer contexts need more memory. A model that downloads successfully may still exceed the available inference capacity. Ollama’s API has no host-published port. Open WebUI reaches it on the private workload network. Ollama has outbound network access for requested downloads.

​
Ask your first question

Create a new chat in Open WebUI, select the downloaded model, and send a question. The model runs on your host. Reopen the chat to check that your conversation was saved.

​
What persists where

Storage is split deliberately: Chats and model weights persist across container replacement. Keep the model name and digest so you can identify it again if its storage is lost. Model availability depends on its upstream source. Complete AI backup and restoration remain unverified; container persistence alone does not establish recovery. See Back up and restore for the workspace lifecycle.

​
Troubleshooting

If a question is not covered above, check the Node Hub status for the workload containers before changing anything in the generated runtime directory.