Set up a local model

Before you start

You need a paired desktop Station on Windows, macOS or Linux, updated to the latest version, and about 8 GB of free disk space. Keep the computer on and connected during setup; downloads resume if interrupted.

Set it up

1. On the Station computer, open the Station menu in the taskbar (the menu bar on a Mac) and choose Allow local model setup. This is the only step you do on that computer.
2. In your workspace, open Settings → Stations. In your Station’s AI provider menu, choose Local model, then Set up a local model.
3. Choose a model. Qwen 3.5 9B is recommended for most cameras.
4. Choose how to get the files. Download fetches the verified files. If Station found them already, for example in LM Studio, choose Use files already on your Station: they stay where they are.
5. Choose where it runs: NVIDIA GPU if Station found one, otherwise the processor (or Apple Silicon on a Mac). Check the notes about memory and disk space.
6. Choose Download and set up. Station downloads the AI runtime if needed, verifies every file and runs a short image-analysis test. You can close the window; setup continues on the Station.
7. When the model is ready, choose Use Qwen 3.5 (or Gemma 4). Cameras on that Station now use it, unless a camera has its own AI provider setting.

Linux

The Linux Station runs as a service without a menu. Allow local setup once from a terminal, using the service’s data folder, then continue in your workspace.

sudo -u agentinosaurus-station /usr/lib/agentinosaurus-station/station \
  -data /var/lib/agentinosaurus-station -allow-local-models

Use files you already have

Station looks in Downloads, LM Studio and the Hugging Face cache. If your files are elsewhere, choose Open local models folder in the Station menu, copy both files there, then choose Look again in setup.

Qwen 3.5 9B needs Qwen3.5-9B-Q4_K_M.gguf and its mmproj-F16.gguf vision file. Gemma 4 12B needs gemma-4-12b-it-qat-q4_0.gguf and mmproj-gemma-4-12b-it-qat-q4_0.gguf. Other versions or quantizations are not accepted.

If something goes wrong

Not enough disk space: free some space on the Station computer and choose Resume setup. Parts already downloaded are kept.

The download stopped: check the internet connection on that computer and resume.

The image test failed: this usually means the computer ran out of memory. Close other apps, try Qwen 3.5, or choose the processor instead of the GPU.

The runtime couldn’t start on the GPU: update the NVIDIA driver, or choose the processor.

The model didn’t start after a restart: Station retries automatically. If the files were moved or changed, set the model up again.

Switch back or remove it

Choose another provider in the Station’s AI provider menu at any time. Disconnect model stops using it but keeps its files, so you can delete them yourself or set the model up again without downloading.