Liquid d1

Liquid d1 is a family of AI models that assigns probabilities to supplied answers. A developer can ask a question about a document or camera image, supply the allowed choices and receive a score for each. This suits tasks such as classifying a request or checking whether an object is visible.

Liquid AI released downloadable d1-3B and d1-omni-600M weights on 7 October 2026. The company had introduced the hosted d1 service with vision support on 5 October. The downloadable models and the hosted service have separate capabilities and access routes.

Two sizes handle different inputs

The d1-3B model accepts text and images. It is built for running on local computers and edge devices, which process data near a camera or other sensor.

The smaller d1-omni-600M checkpoint is experimental. It accepts text with an image or text with audio. Its model card explains the supported input combinations and how to run them.

The output is a decision score

d1 evaluates answer choices in one forward pass through the model. Its output is useful when software needs a bounded decision, such as selecting a category. The quality of the decision depends on the question, answer choices and input.

Both model cards specify the LFM1.0 licence. Teams distributing or using the weights commercially should read that licence and test accuracy on their own inputs. Liquid’s published speed results describe its named hardware and test conditions.