Built on computational units rooted in dynamical systems, signal processing, and numerical linear algebra, Liquid Foundation Models (LFMs) represent a departure from traditional transformer-based architectures12. This innovative approach allows for efficient memory usage and processing of longer data sequences, making LFMs suitable for handling various types of sequential data including text, audio, images, video, and signals13. The models' unique design enables real-time adjustments during inference without the computational overhead associated with traditional models, while maintaining a significantly smaller memory footprint, especially for long-context processing4.

Three distinct models comprise the Liquid AI lineup, each tailored for specific use cases. The LFM-1B, with 1.3 billion parameters, is designed for resource-constrained environments. For edge deployments such as mobile applications, robots, and drones, the LFM-3B offers 3.1 billion parameters. The most powerful model, LFM-40B, is a "mixture of experts" system with 40.3 billion parameters, optimized for complex cloud-based tasks12. These models are currently available for early access through platforms like Liquid Playground, Lambda, and Perplexity Labs, allowing organizations to integrate and test them in various deployment scenarios1.

Early benchmark results indicate impressive performance from Liquid AI's models. The LFM-1B has reportedly outperformed transformer-based models in its size category, particularly excelling in benchmarks like MMLU and ARC-C1. Meanwhile, the LFM-3B has shown competitive results against established models such as Microsoft's Phi-3.5 and Meta's Llama family12. These models demonstrate strengths in general and expert knowledge, mathematics, logical reasoning, and long-context tasks, while currently falling short in areas like zero-shot code tasks and precise numerical calculations3.

Optimization efforts are underway to enhance LFM performance on hardware from major tech companies like NVIDIA, AMD, Apple, Qualcomm, and Cerebras1. Liquid AI has scheduled a full launch event for October 23, 2024, at MIT's Kresge Auditorium, where they plan to showcase their models' capabilities1. Leading up to this event, the company will release a series of technical blog posts detailing the mechanics of each model1. Additionally, Liquid AI is encouraging red-teaming efforts to test the limits of their models and improve future iterations1. While taking an open-science approach by publishing findings and methods, the company will not open-source the models themselves to maintain a competitive edge in the AI landscape2.