Liquid AI releases open-d1, a multimodal open decision model for the edge

On October 7, Liquid AI published open-d1 on the official Hugging Face blog, describing it as a multimodal open decision-model family built for edge devices. Both the channel and the name point in the same direction: a model intended for on-device use and made openly available.

The framing in the post centers on three ideas: multimodal, open, and edge. Multimodal means the model is aimed at more than plain text input; open means the model and its approach are made available to others; edge means it has to run on hardware with tight limits on compute, memory and power.

Moving multimodal capability to the edge has been one of the main threads of on-device AI competition over the past year. Cloud models still lead on depth of understanding and context length, but edge approaches offer more control over latency, privacy and network dependence, which is why chipmakers, device makers and model teams are all investing there.

For Liquid AI, releasing open-d1 openly hands the verification back to developers. The real signal of open-d1 is that it tries to move the combination of multimodal input and decision-making directly onto the device rather than leaving it as a cloud demo.

Public information so far is mostly limited to the blog title and positioning; details such as parameter size, the specific mix of modalities and benchmark results still need to be confirmed against the official release.

What to watch next is whether open-d1 ships with open weights, inference code or examples, and how it performs on common edge hardware. That will decide whether this is a demonstrative release or something that can genuinely enter an on-device product stack.