Startup Liquid AI Redefines AI Landscape With Revolutionary Multimodal Models

CAMBRIDGE, MA — October 1, 2024

Liquid AI, a promising startup co-founded by former Massachusetts Institute of Technology (MIT) researchers, has introduced a groundbreaking suite of multimodal AI models. Strikingly, these models diverge from the widely adopted transformer architecture, spotlighted in the influential 2017 paper "Attention Is All You Need." Instead, Liquid AI seeks to dismantle the reliance on Generative Pre-trained Transformers (GPTs), drawing inspiration from fundamental engineering principles like those used in the creation of engines, cars, and airplanes.

Revolutionary AI Models

Named the "Liquid Foundation Models (LFMs)," these next-generation AI frameworks aim to establish a new benchmark in terms of both performance and efficiency. Three models have been introduced:

  • LFM 1.3B
  • LFM 3B
  • LFM 40B MoE (Mixture-of-Experts)

The "B" denotes the number of parameters in each model, which essentially dictates the model’s capacity for information processing and analysis. Generally, a higher parameter count translates to better model performance across a diverse range of tasks.

Remarkably, the smallest model, the LFM 1.3B, outperforms comparable transformer-based models such as Meta's Llama 3.2-1.2B and Microsoft's Phi-1.5 on several key benchmarks, including the Massive Multitask Language Understanding (MMLU) evaluation, which tests AI across various STEM fields. This achievement represents the first instance where a non-GPT architecture has surpassed traditional transformer-based models in such benchmarks.

Unparalleled Memory Efficiency

One of the standout aspects of the LFMs is their memory efficiency. As Maxime Labonne, Head of Post-Training at Liquid AI, noted on social media platform X, the LFM-3B model requires merely 16 GB of memory during inference, in stark contrast to over 48 GB demanded by Meta's Llama-3.2-3B model. Labonne described the release as the "proudest" moment of his career, emphasizing the models' minimal memory footprint and efficient inference, which facilitate long-context tasks on edge devices for the first time.

A New Approach to AI Architecture

Liquid AI's models boast a unique architecture rooted in the principles of dynamical systems, signal processing, and numerical linear algebra. This approach has yielded general-purpose AI models capable of processing various types of sequential data, including video, audio, text, time series, and signals.

Liquid AI's previous work with Liquid Neural Networks (LNNs) informed much of this new development. Highlighted in an earlier report by VentureBeat, LNNs require fewer neurons than traditional deep learning models to perform complex tasks, enhancing both efficiency and adaptability. The new LFMs maintain these advantages, making real-time adjustments during inference without the heavy computational demands typically associated with traditional models.

Broad Industry Implications

These innovative models are designed to excel not just in raw performance but also in operational efficiency, making them highly suitable for diverse applications across multiple sectors, such as financial services, biotechnology, and consumer electronics. This positions Liquid AI as a significant contender in the AI foundation model space.

Restricted Access and Future Plans

For now, the models are not open source; instead, access is available through Liquid AI’s various platforms, including an inference playground, Lambda Chat, and Perplexity AI. Liquid AI is currently optimising these models for deployment on hardware by NVIDIA, AMD, Apple, Qualcomm, and Cerebras, and invites early adopters and developers to test and provide feedback during this preview phase.

The company plans a full launch event on October 23, 2024, to be held at MIT’s Kresge Auditorium in Cambridge, MA. Interested parties are encouraged to RSVP for the in-person event.

Looking Ahead

In the lead-up to the official launch, Liquid AI promises to release a series of technical blog posts elaborating on the intricacies of their models and methodologies. Additionally, the company is committed to rigorous testing measures, inviting users to push the limits of their models to enhance future iterations.

With the introduction of the Liquid Foundation Models, Liquid AI aspires to reframe how AI models are perceived and utilised, balancing top-tier performance with remarkable efficiency. This strategic move not only challenges existing transformer-based models but also potentially sets new standards for the AI industry.

Source: Noah Wire Services