Both are available on Hugging Face, and Liquid AI says they have day-one llama.cpp support. The company describes the models as producing an answer in one forward pass rather than generating output tokens.
d1-3B accepts text and images. d1-omni-600M accepts text with images or text with audio. Liquid AI reports a 48.57 score for d1-3B on the public split of Decision Index v0.2.1, and 15.95 for d1-omni-600M. It also reports single-question latency for d1-3B of 8 ms on an Nvidia RTX 4090, 16 ms on Jetson AGX Thor, 26 ms on Jetson AGX Orin, and 50 ms on Jetson Orin Nano.
The release says the weights can be downloaded, fine-tuned, and deployed without restrictions, but does not name or reproduce a license in the announcement. The benchmark scores and latency figures are Liquid AI's own measurements. The post does not provide independent validation, report its private vision split, or provide dedicated audio-decision benchmark results.