Local models: camera AI that runs on your own computer

Some rooms should never send a single frame to the cloud. LiquidAI d1-2B now analyzes your cameras on the Station’s own computer, and it’s the default on desktop Stations. No footage leaves your premises for analysis, and there’s no token bill.

Some cameras should never send a frame anywhere

Update, 5 November 2026: Station now sets up the local model by itself after one approval during install or pairing. Read Local-first: camera AI that stays at home for what changed; the steps below describe the earlier dialog.

Update, 4 November 2026: LiquidAI d1-2B is now available on desktop Stations and is their default AI. We’ve updated this article to match.

Think about the camera you hesitated to install. The one in a parent’s living room. The treatment room in a small clinic. The workshop where the prototype sits on the bench. These are exactly the places where an Agent that pays attention is most useful, and exactly the places where sending frames to anyone feels like too much.

Until now, Agentinosaurus had one answer to that worry: your Station sends frames straight to the AI provider you choose, with your own API key, and they never pass through our servers. For most cameras that’s enough. For some places, any third party is one too many.

So we added a third choice next to OpenAI and TypeSafe AI, and made it the default on desktop Stations: a local model. The AI runs on the same computer as your Station, and your camera frames are analyzed right there.

What changes, and what stays the same

With a local model, your Station prepares frames exactly as it does for a cloud provider. Hidden areas are blacked out first, then the model on your computer answers the same question: is any of your Events happening? The frames are never sent to a cloud AI provider, and Station never quietly falls back to one. If the model can’t run, the camera reports a problem instead.

Everything you rely on keeps working. Your workspace still receives the answers, so you get the same Messages, incidents, Push notifications, Webhooks and history. Incident snapshots stay on your Station, as they always have. Your Events, schedules and alerts don’t change when you switch.

Two things to know. Station still needs the internet to receive your Events and deliver your alerts, so this isn’t an offline mode. And Camera Chat, where you ask a camera a question, still uses your OpenAI key.

The model: LiquidAI d1-2B

Station runs LiquidAI d1-2B, a small decision model built on LFM3-VL that answers in one pass. It watches a 20-second observation as a short video at two frames per second, using the same processed frames a cloud provider would get, with hidden areas and blurred faces. The model and its vision file are about 2 GB to download.

We published results on real camera footage in our October benchmark articles. Treat any test as a starting point: try your own Events on your own cameras before you rely on a model.

The model is licensed under the LFM Open License v1.0, which Station shows during setup. Commercial use is free for organizations with less than $10 million in annual revenue; larger organizations need a license from Liquid AI.

Station only runs these exact files. Every download, and every copy you already have, is checked against its published SHA-256 checksum before it’s used.

What your computer needs

Local models work with desktop Stations on Windows, macOS and Linux. Allow at least 8 GB of memory and about 3 GB of free disk space for the model and runtime. An NVIDIA graphics card or a Mac with Apple Silicon makes every check much faster.

In our CUDA test with two Events, the median decision took about 2.4 seconds. Station runs one check at a time, queues a few more, and keeps the model loaded in between. On the processor alone the model still works, just more slowly, which suits one or two cameras that analyze when something moves.

You don’t have to work any of this out in advance. Setup looks at the computer first and tells you what it found: memory, graphics card, free space. It warns you if something looks short. The 40-frame video test at the end is what decides whether the model is ready.

Four steps, and no file paths to type

First, you allow it on the computer itself: open the Station menu and choose Allow local model setup. It’s the only step you do there, and it matters. Our servers can’t start a download or run a program on your computer on their own.

Then the rest happens in your workspace, under Settings → Stations. Choose how to get the files: download them, or use the ones Station already found if you’ve downloaded them before with LM Studio or from Hugging Face. Station reports which model it found, never the folder it found it in. Finally, choose where it runs, on your graphics card or the processor, and start the install.

Station downloads what it needs, verifies every file and runs a short video-analysis test. You can close the window while it works; downloads resume if they’re interrupted. When it says Ready, choose to use the model, and the cameras on that Station start using it. After a restart, Station loads the model again in the background, so your cameras carry on without you.

What it costs

A local model has no AI provider bill. Token usage is still recorded so you can see how busy each camera is, at $0. Your Agentinosaurus plan and camera slots are billed as usual, and the electricity and hardware are yours.

That makes the choice fairly simple. A cloud provider needs no extra hardware or model download, and it works on the Android and iOS apps too. A local model makes sense where frames must stay on your premises, or where cameras watch for many hours a day and you’d rather pay for a graphics card once than for usage every month.

You don’t have to pick one for everything. The local model is the default on desktop Stations; switch individual cameras to a cloud provider where you need one, or the other way round.

Start with the room you’d never put online

You probably already know which camera this is for. The one you never connected, or the one you switched off because it felt like too much. Put a Station on a computer that stays on, set up a local model when it arrives, and give that camera a single Event.

Our step-by-step guide under Learn → Local models walks through every screen, including Linux and what to do if something goes wrong. Your cameras get the attention. Your footage stays at home.