MIT Spinoff Liquid AI Debuts State-of-the-Art Non-Transformer Models
Liquid AI introduces its Liquid Foundation Models, a new line of non-transformer AI architectures that outperform comparable models from Meta and Microsoft while using significantly less memory.
Liquid AI, a startup founded by former MIT researchers, officially launches its first multimodal AI models known as Liquid Foundation Models (LFMs). Unlike the vast majority of modern generative AI systems, these new models completely abandon the standard transformer architecture. Instead, the company builds its models from first principles, comparing the process to how engineers design physical machines like cars and airplanes.
The new LFM family currently includes three sizes: a 1.3 billion parameter model, a 3 billion parameter model, and a massive 40 billion parameter Mixture-of-Experts model. Despite moving away from the dominant GPT framework, these models already achieve state-of-the-art results. The smallest LFM notably outperforms similar-sized models from Meta and Microsoft on major industry benchmarks, marking a significant milestone for non-transformer architectures.
Beyond raw performance, Liquid AI focuses heavily on operational efficiency, particularly regarding memory usage. The 3 billion parameter LFM requires only 16 GB of memory, which is drastically lower than the 48 GB needed by Meta's equivalent Llama model. This combination of high capability and low memory overhead makes the LFMs highly attractive for a wide range of enterprise applications.