Liquid AI's New 'Decision Models' Are What Edge AI Was Missing

The short version

I’m always looking for open-source AI I can run on small devices. So this is exciting. Liquid AI just released two models that don’t chat, they decide—and they’re small enough to run on a device in your hand. This is a huge deal for anyone building AI that needs to work fast, offline, in the real world.

What are these new Liquid AI models?

Let’s get the names out of the way. They’re called d1-3B and d1-omni-600M. The names aren’t catchy, but what they do is. These are not your typical Large Language Models (LLMs). Liquid AI calls them “decision models.” I think that’s a perfect description.

Instead of generating a paragraph of text, they output a structured choice. Think classification, routing, or scoring. Yes or no. This or that. A, B, or C. They give you a direct answer in one shot.

The two models are available right now on Hugging Face. The first, d1-3B, is a 3-billion-parameter model that can understand both text and images. That’s small enough to be interesting for edge devices but big enough to handle complex inputs. The second, d1-omni-600M, is a much smaller 600-million-parameter model. It’s experimental, but it can process text with either images or audio. That audio capability is a very big signal about where this is all going.

These are open-weight models. That means you can download them, run them on your own hardware, and fine-tune them for your specific needs without asking anyone for permission or paying an API fee for every call. For builders, this is everything.

Why is a ‘decision model’ a big deal?

Because most of the AI hype is focused on the wrong tool for the job. Generative models like GPT-4 or Claude 3 are incredible feats of engineering. They write poetry and code. But using one to decide if an email is “urgent” or “not urgent” is like using a sledgehammer to hang a picture frame.

It’s overkill. It’s slow. And it’s expensive.

I spent a good chunk of last month trying to get an open-source LLM to reliably classify customer support tickets into one of five categories. The process is a nightmare of prompt engineering, output parsing, and error handling. You ask the model a question, it gives you a chatty sentence back, and then you have to write more code to find the actual answer buried inside. “Based on the content, I would classify this ticket as: Technical Support.” It’s inefficient.

Decision models skip the whole conversation. They are purpose-built for the task of making a choice. You give them the input, and they give you the classification. Directly. The difference in speed and efficiency is staggering. We’re talking about milliseconds versus multiple seconds. For any application that needs to react in real-time—a factory robot inspecting parts, a smart camera identifying a threat, a medical device monitoring a patient—that difference is the entire product.

This is the core of edge computing. The intelligence needs to live on the device itself, not in a data center thousands of miles away. You can’t afford the latency of a round trip to the cloud when a decision has to be made instantly. Liquid AI’s models are explicitly designed for this world. They are built to be fast and frugal.

How do they actually work?

The key phrase Liquid AI uses is “single forward pass.” It sounds technical, but the concept is simple. The model takes in the data—say, a photo of a plant and the text “Are the leaves yellow?”—and processes it all at once to arrive at a decision. It doesn’t need to generate a bunch of intermediate thoughts or text. It just computes the answer.

This architecture is what makes them so fast and small. A 600 million parameter model is tiny by today’s standards. You can realistically run that on something like an NVIDIA Jetson, which is a small, powerful computer-on-a-module designed for deploying AI on robots, drones, and smart cameras. You could even run it on a high-end smartphone. This is a far cry from the massive, power-hungry server farms needed for giant LLMs.

The multimodality is also critical. The d1-3B model looks at an image and text together. This allows for so much more nuance than a model that can only see one or the other. You can ask it to classify a product review not just based on the star rating, but also based on the photo the user uploaded. Is the product actually broken in the picture? The model can see that.

The d1-omni-600M model goes a step further by adding audio to the mix, paired with text. Imagine a home assistant that can hear you say, “What’s that noise?” and simultaneously analyze the audio file of a beep from the kitchen to tell you it’s the smoke detector’s low-battery warning. That’s the kind of sophisticated, real-time decision-making this enables. You can read the full technical breakdown in Liquid AI’s original announcement post.

What should you do with these models?

If you’re a developer, engineer, or even a curious hobbyist working on anything related to edge AI, you should download one of these models today and try it out. They are on Hugging Face, ready to go. The barrier to entry is low.

Here are a few ideas I’ve been kicking around for my own projects:

1. Smart Inbox Triage: I get a lot of emails with attachments. I could use the d1-3B model to build a local script that classifies incoming emails. It could look at the image attachment and the subject line to decide: Is this an invoice? A family photo? A screenshot for a bug report? It could then automatically file them for me. No cloud API needed, so my data stays private.

2. A Better Security Camera: Instead of just getting a motion alert, a camera running the d1-3B model could give a much more specific notification. It could see a delivery person and a package, and the notification could be “FedEx dropped off a package.” It could see a raccoon, and the notification could be “Raccoon on the porch.” This is much more useful than a generic “motion detected” alert.

3. Real-Time Audio Alerts: Using the tiny d1-omni-600M model, you could build a small, battery-powered device for a workshop. It could listen for the specific sound of a machine malfunctioning or a piece of glass breaking and send an immediate alert. Because it’s running locally, it works even if the Wi-Fi goes down.

4. Interactive Kiosk: A museum could set up a kiosk where you point a camera at an exhibit, and a text prompt asks a question. The d1-omni-600M model could process the image and the text to provide a scored answer or classification on-screen instantly, creating a simple, interactive experience without needing a web connection.

The point is to think beyond the chatbot. These models from Liquid AI push us to build AI that senses, understands, and acts directly within its environment. It’s a move from conversational AI to operational AI.

FAQ

What is a ‘decision model’? It’s a type of AI model designed to output a structured choice, like a category, a score, or a yes/no answer, rather than generating conversational text. They are highly efficient for tasks like classification, scoring, and routing.

Are these models free to use? Yes, they are released as “open-weight” models. This means the model files are publicly available on Hugging Face, and you can use them for both research and commercial projects, according to the terms of their license.

Do I need a powerful computer to run them? No. They are specifically optimized for efficiency and are designed to run on “edge devices” like an NVIDIA Jetson or even high-end mobile phones. The 600M parameter model is especially lightweight.

Can these models replace ChatGPT? No, they are built for entirely different tasks. ChatGPT is a generative model for creating human-like text. Liquid AI’s d1 models are decision models for making fast, structured choices on-device.