Liquid AI Inc. , an artificial intelligence startup and MIT spinoff, has launched its first generative AI models, called "Liquid Foundation Models" (LFMs), which are distinguished by a novel architecture focused on performance and efficiency. Founded by MIT researchers Ramin Hasani, Mathias Lechner, Alexander Amini, and Daniela Rus, Liquid AI aims to develop versatile AI models suitable for all organizational sizes. LFMs utilize principles from dynamical systems, numerical linear algebra, and signal processing, allowing them to efficiently process various sequential data types, such as text and audio. Unlike traditional deep learning models that require thousands of neurons, LFMs, based on Liquid Neural Network (LNN) architecture, achieve comparable outcomes with significantly fewer.

This efficiency allows them to handle up to 1 million tokens with minimal memory usage. The initial offerings include three models: LFM-1B (1. 3 billion parameters for resource-limited environments), LFM-3B (3. 1 billion parameters for edge applications), and LFM-40B (40. 3 billion parameters for complex cloud-based scenarios). These models have already demonstrated superior performance in AI benchmarks, competing effectively against established models like ChatGPT and others. Liquid AI has made the LFMs available for early access through platforms like Liquid Playground, Lambda, and Perplexity Labs, enabling organizations to integrate and test the models in various deployment scenarios. The company is also optimizing the models for specific hardware from companies like Nvidia and AMD, with plans to release detailed technical information and invite the AI community to stress-test the LFMs.