Multimodal open d1 decision models for the edge

LiquidAI Releases Open, Multimodal Decision Models for Edge Computing

LiquidAI has released two open-weight decision models, d1-3B and d1-omni-600M, specifically engineered for high-speed, on-device processing. The release targets the growing demand for efficient AI that can run directly on edge hardware. Notably, the 3-billion parameter d1-3B model achieves a score of 48.57 on the Decision Index 0.2.1 benchmark, outperforming several larger models, including the 35-billion parameter Decider 35B-A3B.

Technical Specifications and Performance

Unlike generative models that produce tokens sequentially, these d1 models are designed to provide structured answers in a single forward pass, making them exceptionally fast for decision-oriented tasks. The models are built on LiquidAI's Liquid Foundation Models and are optimized for multimodal inputs. Performance benchmarks, conducted in collaboration with NVIDIA, highlight their efficiency on the Jetson platform.

• d1-3B: A 3B parameter, decoder-only model supporting text and image inputs. It achieves a mean score of 82.9 on a composite benchmark of seven public datasets.
• d1-omni-600M: An experimental 600M parameter model with a bidirectional encoder, supporting text-and-image or text-and-audio inputs.
• Edge Inference Speed: On an NVIDIA Jetson AGX Thor, d1-3B answers a single question in just 16 milliseconds.
• GPU Inference Speed: On an NVIDIA RTX 4090, a single question is answered in 8 milliseconds.

Impact on the AI Ecosystem

By open-sourcing these high-performance models, LiquidAI provides developers with critical tools for building applications in robotics, IoT, and consumer electronics where low latency and computational efficiency are paramount. This move signals a deliberate focus on practical, deployed AI that operates independently of the cloud, potentially accelerating the integration of sophisticated multimodal understanding into everyday devices. The availability on Hugging Face and provision of clear usage examples further lowers the barrier to adoption for developers working with edge hardware.

LiquidAI is strategically carving out a niche in the AI market by prioritizing specialized 'decision models' over general-purpose generative ones. By open-sourcing models optimized for specific NVIDIA edge hardware, the company is directly addressing the industrial and embedded systems market, moving beyond the cloud-centric generative AI race toward real-world, low-latency applications.