Liquid AI releases Pareto-frontier, multimodal decision model for the edge (www.liquid.ai)

Liquid AI has announced the release of two innovative multimodal decision models: d1-3B and d1-omni-600M, which excel in both speed and performance on edge devices. The d1-3B model, with 3 billion parameters, achieves a remarkable Decision Index score of 48.57, outperforming all other models under 10 billion parameters and rivaling a 35 billion parameter model. It delivers lightning-fast inference times, processing a question in just 8 ms on an NVIDIA RTX 4090 and under 50 ms on various Jetson devices. Meanwhile, d1-omni-600M, an experimental model with 600 million parameters, is capable of handling text, images, and audio, achieving a score of 15.95 on the Decision Index. These advancements are significant for the AI/ML community as they enable real-time, structured decision-making across a range of applications, including those that require multimodal inputs. The models utilize different architectures—d1-3B employs a decoder-only backbone, while d1-omni-600M incorporates both text and audio encoders alongside a visual component. This approach optimizes training and performance, allowing for effective deployment on low-latency edge hardware. Liquid AI encourages the community to explore these models on Hugging Face, enhancing the development of multimodal decision systems.