Liquid AI publishes multimodal open d1 decision-model weights for edge systems

Liquid AI has introduced two downloadable decision-model checkpoints intended for applications that need a structured choice rather than generated text. The larger d1-3B accepts text with images, while the smaller experimental d1-omni-600M supports a text-and-image or text-and-audio combination. In a decision-model workflow, an application supplies a state and a set of typed questions; the system returns option scores or calibrated answers in one forward pass. This makes the release relevant to routing, moderation, visual inspection and agent controls, without implying that either model is a general chat assistant.

The architectures differ. Liquid AI says d1-3B builds on LFM2.5-VL-3B, whereas the smaller experimental model uses its LFM2.5 encoder family with additional modality components. The official d1-3B model card describes 3.12 billion parameters, a 32,768-token context and an LFM 1.0 license. Those details are attached to the d1-3B repository, not automatically transferable to the omni variant. Both releases have Hugging Face model pages, which makes exact file-level inspection possible before the registry assigns verified artifact and license fields.

Liquid AI reports a 48.57 score on Decision Index 0.2.1 for d1-3B and single-question latency of 16 milliseconds on an NVIDIA Jetson AGX Thor. These results are publisher evaluations with their own tasks and hardware setup; Open Model Weights has not independently benchmarked quality or serving speed. The practical significance is a new class of compact, multimodal decision checkpoints that can be assessed separately from generative models. Registry ingestion should keep the two models and their evidence distinct, record source-specific licensing, and avoid treating one task benchmark as a universal ranking.