Liquid AI Releases d1 Models That Decide Without Writing a Word

Liquid AI has published two open-weight models in its d1 family, and the pitch is unusual: neither one writes text back. Instead, they are built to make decisions. d1-3B accepts text and images. d1-omni-600M accepts either text paired with an image or text paired with audio. Both return structured, typed answers from one pass through the network, using zero output tokens.

The distinction matters for speed and cost. A conventional chat model produces its answer one token at a time, and each token adds delay and compute. A model that skips generation entirely can, in principle, respond in about the time it takes to run a single forward pass. That is the basis for the company's stated goal of real-time decisions. Liquid AI also describes the outputs as calibrated, meaning the confidence attached to each answer is meant to track how often the answer is actually correct. That lets downstream software act on the result with a sense of how much to trust it.

The design points toward gating and routing work: the kind of logic that sits between sensors or user inputs and larger systems. Examples include deciding whether a camera frame needs attention, whether a spoken request should trigger a particular action, or whether an input should be sent to a bigger model at all. Those are tasks where a fast, typed yes-or-no style answer is more useful than a paragraph of prose.

Because the weights are open, developers can inspect, run and fine-tune the models on their own hardware instead of calling a hosted API. The small sizes also fit Liquid AI's broader focus on efficient models for on-device and edge deployment. The release materials as summarized here do not include independent benchmark results, so how well the calibration holds up in practice will depend on testing by the wider developer community.