Liquid AI drops non-transformer models that rival top LLMs

The entire AI world has spent the last seven years worshipping the Transformer architecture. Liquid AI, a spin-off from MIT’s CSAIL, thinks it’s time to move on.

The Boston-based startup just unveiled Liquid Foundation Models, or LFMs—a new family of multimodal models that ditch the classic Transformer setup entirely. Instead, they run on "liquid neural networks," drawing on dynamical systems, numerical linear algebra, and signal processing to process sequential data with drastically less compute.

Squeezing big context into tiny memory

Liquid rolled out three models covering everything from drones to cloud servers.

At the low end is LFM-1B, a 1.3-billion-parameter dense model built for tight hardware constraints. Liquid claims it outpaces Meta’s Llama 3.2-1.2B and Microsoft’s Phi-1.5 on benchmarks like MMLU and ARC-C, setting a fresh standard for the 1B weight class.

Next is LFM-3B, packing 3.1 billion parameters for mobile phones, robots, and drones. Its standout trick is memory management: while long inputs typically blow up GPU memory, LFM-3B handles long contexts with just 16 GB of memory, compared to over 48 GB required by Meta’s Llama 3.2-3B.

At the top sits LFM-40B, a 40.3-billion-parameter Mixture-of-Experts model designed for complex enterprise workloads. The company claims these architectures can process up to 1 million tokens efficiently while adapting in real time during inference.

Liquid AI’s head of post-training, Maxime Labonne, called the launch the proudest release of his career on X, highlighting that the models beat Transformer-sized rivals while slashing memory usage. But unlike Llama, LFMs are not open-source—you have to test them through Liquid's playground, Lambda Chat, or Perplexity AI.

Why it matters

For builders running workloads at the edge, LFMs are a massive deal. If you can analyze long technical manuals or stream sensor data on a 16 GB edge device without offloading to a cloud cluster, hardware bills drop dramatically. It is definitely worth running some test prompts on Perplexity or Lambda Chat to see if the reasoning holds up to your current setup.

For the broader industry, Liquid AI is proving that raw brute-force scaling isn't the only way forward. Transformers are notorious memory hogs, and hardware vendors have extracted huge premiums just to keep up with their VRAM demands. An architecture that solves this mathematically could force competitors to rethink alternative designs.

That said, we think Liquid AI faces an uphill battle with adoption. Skipping the open-weights route leaves them without the grassroots army that turned Llama into the default enterprise playground. Competing against established giants with a proprietary, non-standard architecture means their efficiency gains have to be flawless to convince risk-averse enterprises to switch.