What changed
Liquid AI released d1-3B and the experimental d1-omni-600M on October 7. The company describes them as decision models: they answer in a single forward pass rather than generating a stream of words. The larger model accepts text and images; the smaller research release supports text with images or audio. Liquid AI reports benchmark and latency results, while explicitly noting that it does not provide vision or audio benchmark results in this release. Treat its published performance numbers as vendor measurements with specific test conditions.
Source: Liquid AI: Multimodal open d1 decision models for the edge ↗
Our take: match the tool to the job
Imagine a small app deciding whether an uploaded picture contains a receipt or whether a message belongs in the support queue. A long explanation can be unnecessary overhead when the application only needs a category and a confidence threshold. Our reading of this direction is that useful AI products will increasingly combine different kinds of models. One component can route a request; another can write a response; an ordinary rule can decide whether a human must approve it. A chat window is only one possible interface.
A better beginner project
Our suggestion is to prototype a three-category sorter before attempting an autonomous agent. Use harmless sample data, keep an obvious manual correction button and log uncertain predictions. Decide what happens when the system is wrong: a missed study-note label is inconvenient, while an incorrect safety decision could be serious. Fast outputs are appealing, but the useful metric is the number of mistakes your workflow can tolerate. Speed alone does not answer that question or remove the need for testing.
Try this, then make it yours.
List three yes/no or category decisions in an app idea and identify which ones should always allow human correction.
Explore the tool ↗Follow the signal.
Our reporting starts here. Practical suggestions are our analysis, and vendor performance statements are claims unless independently verified. We haven’t hands-on tested this release.
