Liquid AI Releases Open Multimodal d1 Decision Models for Edge Inference

Liquid AI has released two open-weight decision models, d1-3B and d1-omni-600M, engineered for rapid, structured inference at the edge. Unlike conventional generative transformers that produce tokens sequentially, these models are built on Liquid Foundation Models and deliver deterministic answers in a single forward pass. Available immediately on Hugging Face, the release targets applications requiring fast, reliable decision-making across text, vision, and audio modalities without the latency of autoregressive decoding. Benchmark evaluations across seven public datasets spanning reading comprehension, toxicity detection, intent classification, medical QA, and cross-lingual understanding demonstrate significant efficiency gains. The d1-3B model achieves a mean score of 82.9, outperforming the Decider 4B baseline while maintaining a smaller footprint. The experimental d1-omni-600M variant reaches a mean score of 78.4, surpassing the Decider 2B model despite utilizing only a quarter of its parameters. Both architectures support multimodal processing, with d1-3B preserving vision understanding from its LFM2.5-VL-3B backbone and d1-omni-600M natively handling text, vision, and audio inputs. Performance testing conducted alongside NVIDIA highlights the models suitability for diverse hardware environments. On desktop GPUs, including the NVIDIA GeForce RTX 4090 and AMD MI325X, d1-3B resolves single queries in under 10 milliseconds and processes standard 384-pixel images in under 18 milliseconds. Edge deployment delivers comparable responsiveness, with the model answering queries in under 50 milliseconds across the NVIDIA Jetson AGX Thor, Jetson AGX Orin 64 GB, and Jetson Orin Nano. Throughput scales efficiently for batched workloads, with the Jetson AGX Thor sustaining 262 decisions per second under packed state conditions. Liquid AI has open-sourced both architectures to streamline integration into resource-constrained production systems. Deployment requires the Transformers library version 5.14 or higher, with model code loaded directly via the trust_remote_code parameter. Developers can configure structured output schemas, including null validation, categorical classification, and numeric scoring, enabling direct mapping of model outputs to application logic without post-processing. The d1-3B model is optimized for environments prioritizing decision accuracy, while d1-omni-600M is tailored for deployments with strict memory or compute constraints. This release underscores a broader industry shift toward specialized, single-pass inference architectures that eliminate computational waste in real-time AI pipelines. By bypassing token-by-token generation, Liquid AI positions these models for low-latency edge deployment, industrial automation, and high-throughput customer interaction systems. Full model weights, integration documentation, and usage examples are publicly accessible for immediate developer adoption.