AI models
Local models
Run camera AI on your own computer: frames stay on your premises and there is no token bill.
What a local model does
With a local model, your Station analyses camera frames on the same computer it runs on. The frames are masked and prepared exactly as for cloud providers, then answered by the model in place. They are never sent to a cloud AI provider.
Your workspace still receives the answers, so you get the same Messages, incidents, Push notifications, Webhooks and history. Snapshots stay on the Station, as always.
Models
LiquidAI d1-2B Q8: a compact decision model. Station analyzes a 20-second observation at 2 FPS (40 frames), using the same processed frames as cloud providers.
No API key is needed. Other local model families are not offered.
Every camera starts with Needs attention, a built-in Event for visible signs of danger. It saves incidents and Messages, with Push, Email, Webhooks and Zapier off by default. Enable individual channels or detach the Event in camera settings. Your own Events are evaluated independently.
Station only runs these exact, verified files. Every download and every file you already have is checked against its published SHA-256 checksum before use.
What your computer needs
A desktop Station on Windows (x64), macOS (Intel or Apple Silicon) or Linux (x64 or ARM64). The Android and iPhone apps use cloud providers.
Allow at least 8 GB of system memory. GPU acceleration can greatly reduce latency; capacity depends on the device, the number of cameras and the Events evaluated.
The model and vision file total about 2.05 GB, plus the runtime, download cache and temporary working space. Setup reports the disk space required.
Setup shows the available resources. A 40-frame video test must pass before Station reports the model as ready.
Speed and capacity
A recent Q8 CUDA replay with two Events had a median decision latency of about 2.39 seconds. Older blog benchmarks explicitly labelled bfloat16 used different weights and are not performance or accuracy claims for this Q8 release. Test your own camera workload.
How many cameras one computer can handle depends on its GPU, how often each camera analyses and how many Events it watches for. Movement-based analysis needs far fewer checks than continuous analysis.
Costs
A local model has no AI provider charge: token usage is still recorded, at $0. Your Agentinosaurus plan and camera slots are billed as usual, and the computer’s electricity and hardware are yours.
Privacy and safety
Model downloads start only after you allow local setup on the Station itself, from its menu. Our servers can’t start a download or run a program on your computer on their own.
The model runs on your computer only, on a private local port with a key that changes each start. Station never falls back to a cloud provider: if the model can’t run, the camera reports a problem instead.
Station looks for existing model files in Downloads, LM Studio and the Hugging Face cache, and reports only which model it found, never the folder.
Questions & answers
Do camera frames leave my computer with a local model?
Not for detection. Frames are analysed on the Station computer. Answers, incident details and device health still reach your workspace. Camera Chat is separate and uses OpenAI.
Can I use a different model or my own files?
Only the supported, verified model files. If you already downloaded them, for example with LM Studio, Station uses them where they are after checking them.
What happens if the computer restarts?
Station loads the model again in the background after it starts, so cameras resume without you doing anything.