The AI does not load
The AI service is a separate process from the agent, with its own way of failing. Work down this list — the order reflects how often each one turns out to be the answer.
1. Is the AI service actually running?
The agent and the AI service are different things. The agent can be online and serving files while the AI service has never started.
Start it from the AI Hub in the configurator, or from the AI screen in the browser. The status says which of "stopped", "starting", "loading weights" and "online" you are looking at — loading weights is not stuck, it is working.
AiAutoStart decides whether the service comes up with the agent; it is off by default.
2. Is the model downloaded?
A model that has been chosen is not a model that has been downloaded. The panel says "Model not downloaded" plainly, and the fix is one button.
Downloads are multi-gigabyte and go to the device, not your browser. A cancelled or interrupted download leaves nothing usable; start it again.
3. Is there enough memory?
This is the most common real failure, and the configurator warns before it happens: "model needs ~7 GB, only 4 GB available".
Rules of thumb:
- A 1–2 B model needs a couple of gigabytes free.
- A 7–8 B model needs eight or more.
- Loading a specialist model adds to that, for as long as it stays loaded.
When memory runs short the machine starts swapping, and the symptom is not a clean error: the service becomes unresponsive, status checks time out, and the AI looks "off" while the disk grinds. If that is what you are seeing, pick a smaller model rather than waiting.
On a Raspberry Pi with less than 8 GB, install with --no-ai and do not run models there at
all.
4. Are the Python dependencies installed?
The first AI start installs them, with a progress window. If that step failed — no network, no disk space, no Python — the service cannot come up.
- Windows, installer — a private Python is provided, so this is rare.
- Windows, portable — you install Python yourself and run
setup_ai.bat; see running Filemote without installing it. - Linux/macOS — the install script handles it, unless you passed
--no-ai.
In the AI Hub, the Environment step of the wizard re-runs this check on demand.
5. Has something else taken port 8765?
The AI service listens on 8765 on the machine itself. If another program holds that port, the service cannot bind and fails at startup.
Related, and worth knowing because it looked like a bug for a while: the configurator and the agent can both own the same AI service. There was a version where Stop ended only the handle the window held, so a service started by the other side survived and a status poller brought it back — the AI appeared to start itself. That is fixed; the AI Hub has the detail.
6. Is the model actually a chat model?
Not every model on a model hub can hold a conversation. The configurator validates before
loading and says so: not found, or "not suitable for chat". Models with restricted access also
need a Hugging Face token, which the install scripts accept as --hf-token.
Symptoms that mean something specific
| What you see | What it usually is |
|---|---|
| Chat says the agent is not loaded | Service running, no model loaded |
| First answer very slow, then fast | Normal: weights loading into memory |
| Empty replies | Reply-length cap reached — AiMaxNewTokens |
| Status times out, machine sluggish | Swapping: model too large for this machine |
| Knowledge base shows counts but no controls | Service stopped; the figures come from disk |
If none of that is it
The AI service keeps its own log, separate from the agent's — see reading the logs. The AI Hub also shows startup errors as they happen, which is usually faster than reading a file.
Updated 2026-08-06