Liquid AI releases open-weight d1 decision models for edge devices
The downloadable d1-3B and experimental d1-omni-600M make bounded decisions from multimodal inputs instead of writing free-form answers.
Bounded decisions from multiple input types
Liquid AI published two downloadable members of its d1 decision-model family on October 7: d1-3B and an experimental d1-omni-600M. Instead of generating a sequence of words, the models answer developer-specified questions in a single forward pass. The 3B-parameter model accepts text and images. The smaller model accepts text paired with images or audio, but Liquid AI describes it as an early research release. The Hugging Face repositories are public and list a custom LFM 1.0 license, so availability of the weights should not be confused with an unrestricted open-source license.
Company benchmarks have limits
Liquid AI reports a 48.57 score for d1-3B on Decision Index 0.2.1, ahead of the models it compares in the under-10B category. It says a single question takes 16 milliseconds on an Nvidia Jetson AGX Thor and 50 milliseconds on a Jetson Orin Nano in its tests. These figures come from Liquid AI’s evaluation, not an independent deployment study. The company did not report vision or audio benchmark scores in this release, and it did not publish speed measurements for the experimental 600M model. Developers considering routing, classification or agent guardrails still need to test error rates, latency and license terms for their own workloads.