Liquid AI

Liquid AI positions itself around a simple premise: advanced intelligence should run on the processors outside data centers. Its site presents Liquid Foundation Models as device-native systems built for the latency, privacy, and hardware limits of the physical world, spanning phones, laptops, cars, and other edge environments rather than only cloud endpoints.

The core offering is a broad model family. The site reports 60 LFMs shipped, over 54 million downloads, and more than 5,100 variants, including LFM2.5 releases for local agents, vision-language models that can see screens, and encoder models that run on CPU. A dedicated decision model, d1, is highlighted for sub-50ms decisions, and the models run on llama.cpp, MLX, ONNX, and CoreML.

Beyond model weights, the site covers research and engineering writing on topics such as agent loops at scale and fixing repetitive “doom loops,” plus solution pages for deployments in automotive, space, e-commerce, financial services, bio, and defense. A Mercedes case study illustrates private in-car AI, reinforcing the privacy-first, data-stays-on-device message.

This makes the site most relevant to ML engineers, embedded and edge developers, and product teams in regulated or hardware-constrained industries who need to run models locally. Researchers tracking efficient architectures and runtime support will also find the blog and model listings useful.

A sensible approach is to start from the model catalog and runtime compatibility list