Liquid AI has released open weights for two models that return classifications or scores in one processing pass. Developers can use them to route customer requests, assess refunds or rank urgency. The larger model accepts text and images; the smaller, experimental model adds an audio path. The release gives access to specialized decision tools while audio evaluation remains unfinished.
Liquid AI released two open-weight decision models on October 7: d1-3B and the experimental d1-omni-600M.[1], [2]
They take a supplied question and return structured decisions in one forward pass. A result can be a yes-or-no decision, a choice among options or a numerical score. The models calculate associated probabilities in that same pass, without first writing prose for another program to parse.[1]
The larger d1-3B accepts text and images and is based on Liquid’s LFM2.5-VL-3B model. The smaller d1-omni-600M combines text with either images or audio. Liquid describes the latter as an early research release. Both sets of weights are available on Hugging Face, a platform for sharing models. The release provides downloadable weights; that access does not establish that the entire development process is open source.[1]
Liquid’s examples include assigning a customer request to billing or technical support, assessing fraud, deciding on a refund and scoring urgency. Multiple inputs can be processed together without padding them all to the same length. The company’s text evaluations cover public datasets, while vision and audio benchmark results are absent from this release. Liquid identifies audio evaluation as an open research problem and offers demonstrations through System One Arcade.[1]
