Liquid AI d1: open decision models that answer in one forward pass (huggingface.co)

Liquid AI has announced the d1 decision models, which deliver highly efficient and accurate responses using a single forward pass. The d1-3B model, boasting under 10 billion parameters, outperforms its peers—scoring 48.57 on the Decision Index 0.2.1 and besting larger models, including Decider 35B-A3B. Supporting multimodal inputs, d1-3B excels with text and images, while the d1-omni-600M model integrates text, images, and audio inputs, showcasing a significant leap in processing capabilities, achieving response times as fast as 16 ms on advanced NVIDIA hardware. The significance of these developments lies in their potential to enhance decision-making frameworks across various applications in AI and machine learning. Both models are built on Liquid Foundation Models, with d1-3B utilizing a decoder-only architecture for high throughput, while d1-omni-600M employs a bidirectional encoder to handle diverse inputs efficiently. The impressive benchmark results highlight the models' effectiveness, with d1-3B achieving an average score of 82.9 across multiple tasks, including reading comprehension and medical Q&A. This release opens the door for new applications in real-time decision-making, harnessing flexibility and speed for practical AI solutions.