1.9B Model Sets Its Sights on Restaurant Cameras: Zero-Token Generation, Runs on a Laptop

A 1.9B-parameter model watches restaurant cameras on a laptop and tells servers in real time which tables need to be cleaned. No cloud, no token generation. This is a weekend project by Roman Khrupa using Liquid AI's d1-2B model.

Notice one detail: it generates zero tokens. That means it is not the kind of language model where you ask a question and it answers. It directly outputs a structured judgment—which table needs cleaning. No text generation, no autoregressive decoding. The model looks at the image, gives an action recommendation, and that's it.

This is somewhat different from the current narrative of large models competing on parameter count and context length. A 1.9B model does one specific job, runs locally, doesn't rent GPUs, and doesn't call APIs. The cost is negligible, and privacy isn't a concern—the video stream never leaves the device.

Liquid AI is a company from MIT, working on so-called 'liquid neural networks.' d1-2B is one of their vision models. Roman hooked it up to a camera over the weekend, wrote a simple interface, and turned it into a usable tool. No training, no fine-tuning, just direct inference.

Some might say this is just a demo. But think about it: restaurant cameras are already there, and servers' phones can receive notifications. What's missing isn't hardware; it's the layer that translates visual signals into action instructions. Previously, that layer either relied on a person watching monitors or on cloud-based large models, with high latency, high cost, and significant privacy risks. Now a model under 2B can do this on edge devices.

Of course, it can only tell you which table needs to be cleared. How to clear it, which table to clear first, and whether the guests are still there are another matter. But that's the logic of small models: they don't aim for generality, they solve one specific problem. And they solve it cheaply enough to be deployed anywhere.

In the video, you can see the model boxes the tables that need cleaning, and the interface is simple and direct. No fancy chat box, no 'Hello, I'm an AI assistant.' It's just an eye plus a judgment.

The project link is in Roman's tweet. If you have a development background, you can look at the d1-2B documentation. If not, just remember the trend: models are getting smaller, more specialized, and cheaper. Those large-model companies that survive on API call fees may need to think about what their next story will be.