Answers your typed questions about a live camera frame, a drawing or a message, fast enough to play games with

You give d1 something to look at (a webcam frame, a sketch, a photo, a support ticket, or a mix of text and images) and a short list of questions with fixed answer shapes: yes or no, a pick from options you name, or a rating on a scale you describe. It answers all of them in one go and puts a probability on every option. It never writes a sentence, so there is nothing to parse and no invented detail to check.

What is new compared with Clef is the size and the speed. At 3 billion parameters it answers a single question in about 8 milliseconds on a gaming GPU and 30 on a recent Mac, quick enough to ask the same question of every frame coming off a camera. Liquid built ten small games around that: open your palm to make a bird fly, freeze when the screen says red light, count your squats, draw something and let it guess. Each game is just typed questions asked again and again as the frames come in.

The System One Arcade Space runs in your browser and uses your webcam. Start with Red Light, Green Light: move on green, freeze on red, and watch how quickly it catches you moving. Then open Quick Draw and sketch something from its list of sixteen words to see the confidence bars shift as you add lines. It runs on a shared Hugging Face GPU, so expect a queue at busy times. A separate Rock, Paper, Scissors demo runs entirely inside Chrome on your own graphics card after a one time 3.8 GB download, so the camera feed never leaves your machine. Running the model yourself is a pip install and a dozen lines of Python.

It judges at a glance and does not reason, so keep the questions to things a person could answer instantly from what is in front of them. The license allows free commercial use only below 10 million dollars in annual revenue; larger companies need a separate agreement with Liquid AI.

• Turn a webcam into a sensor: is anyone at the desk, is the door open, is a helmet on.
• Count reps or check form in a home workout app.
• Route incoming support messages by team and urgency, with low confidence cases sent to a person.
• Check product photos against a checklist before they go live.
• Build gesture control into a prototype without training anything.

Try the demo

View model page

Read this issue on neuralexpedition.com →