Liquid AI, an artificial intelligence startup based in Massachusetts, has unveiled its inaugural generative AI models that do not rely on the conventional transformer architecture. Known as the Liquid Foundation Model (LFM), this new architecture marks a departure from Generative Pre-trained Transformers (GPTs), which underpin well-known AI frameworks including OpenAI’s GPT series, Gemini, and Copilot. The company asserts that these new AI models, constructed from fundamental principles, exceed the performance of large language models (LLMs) within a similar size category.
Introduction of Liquid Foundation Models by Liquid AI
Co-founded by researchers from the Massachusetts Institute of Technology’s (MIT) Computer Science and Artificial Intelligence Laboratory (CSAIL) in 2023, Liquid AI aims to develop advanced architecture for AI models that can match or even outpace current GPT standards.
The LFMs are available in three parameter sizes: 1.3 billion, 3.1 billion, and 40.3 billion. The largest of these is a Mixture of Experts (MoE) model, which comprises several smaller language models designed to address more complex challenges. These models can now be accessed through the company’s Liquid Playground, Lambda for Chat UI and API, and Perplexity Labs, with plans to integrate them into Cerebras Inference soon. Liquid AI further notes that the models are being optimized for compatibility with hardware from Nvidia, AMD, Qualcomm, Cerebras, and Apple.
LFMs are distinct from GPT technology, built on the foundational concept of first principles. This method deconstructs complex technology to its basic components, allowing for a comprehensive rebuild from the ground up.
The startup explains that the new models are constructed utilizing a system of computational units, which represents a rethinking of the traditional token model. Labeled as the Liquid system, these units encapsulate condensed information aimed at enhancing knowledge capacity and reasoning abilities. Liquid AI maintains that this novel approach contributes to lowered memory expenses during inference while boosting performance across multiple domains including video, audio, text, time series, and signals.
Moreover, the company claims that the flexible architecture of the Liquid-based models allows for automatic optimization tailored to specific platforms, depending on their unique requirements and inference cache sizes.
Although the startup’s assertions are ambitious, true performance and efficiency metrics will emerge as developers and businesses adopt these models in their AI processes. However, the company has not disclosed the origins of its datasets or any safety mechanisms embedded within the AI models.
