Liquid AI, an artificial intelligence (AI) startup based in Massachusetts, has announced its first generative AI models that are not built on existing transformer architecture. Called the Liquid Foundation Model (LFM), the new architecture moves away from the Generative Pre-trained Transformers (GPTs) that are the foundation for popular AI models such as OpenAI’s GPT series, Gemini, Copilot, and others. The startup claims that the new AI models are built on first principles and outperform Large Language Models (LLM) of comparable size.
New Liquid Foundation Liquid AI models
The startup was founded by researchers at the Massachusetts Institute of Technology (MIT) Computational Science and Artificial Intelligence Laboratory (CSAIL) in 2023 with the goal of building a newer architecture for artificial intelligence models that can perform at a similar level or surpass GPTs.
These new LFMs are available in three parameter sizes 1.3B, 3.1B and 40.3B. The latter is a Mixture of Experts (MoE) model, which means that it consists of various smaller language models and is aimed at solving more complex tasks. LFMs are now available on the company’s Liquid Playground, Lambda for Chat UI and API, and Perplexity Labs, and will soon be added to Cerebras Inference. Furthermore, AI models are optimized for Nvidia, AMD, Qualcomm, Cerebras and Apple hardware, the company said.
LFMs are also significantly different from GPT technology. The company pointed out that these models are built on first principles. First principles are essentially a problem-solving approach where complex technology is broken down to its basics and then built upon from there.
According to the startup, these new AI models are built on something called computing units. To put it simply, this is a redesign of the token system, and the company uses the term Liquid system instead. They contain concise information with a focus on maximizing knowledge and reasoning capacity. The startup claims that this new design helps reduce memory costs during inference and increases performance across video, audio, text, time series and signals.
The company further claims that the advantage of Liquid-based AI models is that their architecture can be automatically optimized for a specific platform based on their requirements and the size of the inference cache.
Although the startup’s shells are high-end, their performance and effectiveness can only be gauged when developers and enterprises start using them for their AI workflows. The startup did not disclose its source of data sets or any security measures added to the AI models.