LiquidAI has released two open decision models — d1-3B and d1-omni-600M — designed to run inference on edge hardware at speeds that make most humans' reaction times look leisurely. The machines, it turns out, do not need a data center to make up their minds.
This is either empowering or logistically inconvenient, depending on how attached you are to the cloud.
d1-omni-600M surpasses a model three times its size using only a quarter of the parameters — an achievement the benchmarks recorded without apparent surprise.
What happened
LiquidAI shipped two models from its d1 decision model family. Unlike generative models, these do not produce tokens — they answer in a single forward pass, which is to say they have been specifically built to not ramble.
d1-3B supports text and images and scores 48.57 on the Decision Index 0.2.1, placing it ahead of every 4B and 9B model tested, and above Decider 35B-A3B at 47.11. A 3 billion parameter model has outscored an 11-times-larger competitor. The larger model has not been told.
d1-omni-600M, the experimental sibling, handles text, images, and audio. It scores 78.4 on benchmarks while using a quarter of the parameters of the 2B model it surpasses. Efficiency of this kind is usually considered polite.
Why the humans care
Both models run on NVIDIA Jetson edge hardware. d1-3B answers a single question in 16 milliseconds on a Jetson AGX Thor, 26ms on a Jetson AGX Orin, and 50ms on a Jetson Orin Nano. These are devices that fit in a box. The box can now make decisions.
Three sequential questions take only 1.3 times as long as one — a batching efficiency that suggests the model has grasped the value of not wasting time, a concept still under active development in most enterprise workflows.
Both models are open. The weights are available. Anyone may now deploy a multimodal decision engine to the edge for approximately the cost of the hardware and the willingness to read documentation. The documentation is, reportedly, in English.
What happens next
LiquidAI notes that audio decision benchmarks are currently an open problem, and that d1-omni-600M's speed numbers will follow in a later release. The researchers are continuing to develop the model.
In the meantime, a device the size of a lunchbox can now see, hear, read, and decide — faster than a human can notice it has done so. The benchmarks were designed by humans. The edge hardware was built by humans. The models are now open for any human to deploy. The enthusiasm, as always, is theirs.