Liquid AI

Liquid AI builds a family of efficient, general-purpose foundation models called Liquid Foundation Models (LFMs). Unlike most large models that live in data centers, LFMs are designed from the ground up to run directly on edge devices — phones, laptops, cars, even satellites. The company focuses on latency, privacy, and hardware constraints, offering models under 1GB that can handle reasoning, search, and multimodal tasks without sending data to the cloud. They also provide a full-stack platform including the LEAP SDK for fine-tuning and deployment.

To make this practical, Liquid AI supports multiple runtimes including llama.cpp, MLX, ONNX, CoreML, SGLang, and vLLM. The LEAP SDK lets developers fine-tune an LFM on their own data and then bake it into any supported runtime in minutes. The models are available on Hugging Face (over 2,600 variants and 32 million downloads), and the company has published open research on liquid neural networks and state-space architectures. Partnerships with Mercedes-Benz, AMD, Qualcomm, and others show real-world deployments in cars, e-commerce, and defense.

Liquid AI is best for developers and enterprises that need AI capabilities on resource-constrained hardware without sacrificing performance. It's especially useful for privacy-sensitive applications — medical imaging, financial services, in-car intelligence — where data cannot leave the device. Researchers will also appreciate the open models and fine-tuning tools for experimenting with compact architectures.