Liquid AI's October 7 announcement releases two models in its open d1 decision-model family: d1-3B and d1-omni-600M. The smaller model is labeled experimental and described as an early research release under further development. This is a model-release story, not proof of reliability in every application.
Unlike a generative chat model, these models answer named questions in a single forward pass without producing output tokens. The model cards describe typed answers for decisions such as yes/no judgments, choices among options and ratings. They are not models that write a conversational response.
The company says d1-3B accepts text and images and is built from its LFM2.5-VL-3B backbone. The smaller d1-omni-600M uses a bidirectional-encoder backbone with vision and audio encoders. Its inputs can pair text with images or text with audio; the model card says a request cannot carry images and audio together.
Liquid AI reports mean scores of 82.9 for d1-3B and 78.4 for d1-omni-600M across seven public text datasets. Those are company-reported aggregate benchmark results, not independent testing or a guarantee on each task. A mean score should not be confused with a particular application's error rate.
The release reports d1-3B single-question timings of 16 milliseconds on Jetson AGX Thor, 26 milliseconds on Jetson AGX Orin 64 GB and 50 milliseconds on Jetson Orin Nano. These timings refer to the stated test configuration and workload, not an end-to-end latency promise for arbitrary inputs. The company explicitly does not report speed numbers for the smaller experimental model.
The announcement says it does not report vision or audio decision benchmarks in that post; model cards provide additional technical detail. Developers should keep the source, benchmark version, input type and workload attached to each claim. The release does not establish clinical safety, production suitability or a reason to trust decisions without application-specific evaluation.