Multimodal open d1 decision models for the edge

Hugging Face introduces two new open decision models, d1-3B and d1-omni-600M, as part of its d1 decision model family. Unlike generative models, these models provide answers in a single forward pass without token generation. Built on Liquid Foundation Models, they target edge deployment where latency and efficiency are critical. The release marks a shift toward structured decision-making in constrained environments.

Benchmarking on seven public datasets covering reading comprehension, toxicity detection, intent classification, medical QA, and cross-lingual understanding shows d1-3B achieves a mean score of 82.9, outperforming Decider 4B. The smaller d1-omni-600M scores 78.4, exceeding Decider 2B’s 77.1 with only a quarter of the parameters. Vision capabilities of d1-3B are validated on standard benchmarks, while d1-omni-600M supports all three modalities, though audio benchmarks remain an open challenge.

Collaboration with NVIDIA tested d1-3B on multiple edge devices, including NVIDIA GeForce RTX 4090, Jetson AGX Thor, Jetson AGX Orin 64 GB, and Jetson Orin Nano. Inference times remain under 50 ms per question across all devices, with minimal overhead for multiple queries. On GPU platforms, d1-3B answers a question in under 10 ms and processes a 384px image in under 18 ms, demonstrating strong performance in low-latency scenarios.

The models are positioned for applications requiring fast, structured decisions, including multimodal inputs. d1-3B offers the highest decision quality at its size, while d1-omni-600M is designed for environments where model footprint is a priority. Instructions for running d1-omni-600M are available in its model card, with limited examples provided for brevity.