Liquid AI Inc., an AI startup and MIT spinoff, has launched its first generative AI models known as Liquid Foundation Models (LFMs), which aim to transform the AI landscape with a new architecture.

Founded by researchers from MIT, including Ramin Hasani and Daniela Rus, Liquid AI is focused on creating efficient general-purpose AI models for organizations of all sizes.

The LFMs are available on platforms such as Liquid Playground, Lambda, and Perplexity Labs, with plans for future availability on Cerebras Inference.

Designed for efficiency, LFMs utilize minimal memory while delivering high computing power, enabling them to handle various sequential data types including text, video, and audio.

These models are claimed to deliver performance comparable to or superior to existing large language models (LLMs) like OpenAI's GPT and Google's Gemini.

The LFM-3B model has been recognized for its exceptional performance, surpassing both smaller and previous generation models while being more efficient.

Liquid AI's LFM 1.3B model reportedly outperforms Meta's Llama AI and Microsoft's Phi model based on leading AI benchmarking systems.

Unlike traditional models such as ChatGPT, LFMs utilize a redefined token system called the Liquid system, which maximizes knowledge capacity and reasoning while reducing memory costs.

The LFMs can operate on smaller memory chips, enabling their use in portable devices and applications at the 'internet of things' level, such as drones and robots.

The launch of LFMs represents a significant advancement in AI, with the potential to reshape enterprise AI integration and establish a new era of artificial intelligence.

Liquid AI commits to transparency, planning to publish findings and methods while keeping the models proprietary for now due to significant resource investment.

Despite promising benchmarks, Liquid AI acknowledges that the technology is still in early stages, and real-world effectiveness and scalability remain to be tested.