Liquid AI

Liquid AI builds Liquid Foundation Models (LFMs), small and efficient open-weight models designed to run locally on phones, laptops, and edge hardware without a cloud connection. The current LFM2 and LFM2.5 families include vision and audio variants and can run on hardware as small as a Raspberry Pi. Backed by a $250M AMD-led Series A, Liquid targets developers and enterprises building private, offline, low-latency AI, though the models are small and not frontier-class.

Liquid AI, spun out of MIT CSAIL, builds Liquid Foundation Models (LFMs): a class of small, memory-efficient generative models that use structured, adaptive operators rather than a pure transformer stack, optimized for fast on-device and edge inference. The pitch is private, low-latency AI that runs on constrained hardware, from phones and laptops to IoT and automotive systems, without a cloud round trip. The current generation, LFM2 and LFM2.5, spans small sizes from a few hundred million parameters up to around 2.6 billion, with multimodal variants for vision (LFM2.5-VL) and audio (LFM2.5-Audio) and a dedicated Japanese variant. Liquid publishes open weights on Hugging Face and reports strong efficiency benchmarks, though these are largely vendor-published and independent validation is limited. Models as small as roughly 300MB can run on devices with modest RAM, and the LFM2.5-2.6B model has been highlighted running on Raspberry Pi-class hardware. For developers, Liquid ships the LEAP SDK (Kotlin Multiplatform, covering iOS, Android, desktop, and more) and the Liquid Apollo app for fully local, private chat. Open weights are free to download, LEAP is free under a developer license, and premium enterprise features are planned under a separate commercial license handled via sales inquiry. Liquid is well-capitalized, having raised a $250M Series A led by AMD in December 2024, and targets teams building privacy-sensitive, offline, or edge AI rather than those needing frontier-scale reasoning.