Liquid AI, the maker of LFM models for on-device AI

An MIT spin-off

Liquid AI was founded in 2023 by Ramin Hasani, Mathias Lechner, Alexander Amini and Daniela Rus, who came out of MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL). The company, Liquid AI, Inc., does not name its headquarters city on its official pages. Its stated mission is to build efficient general-purpose AI “at every scale,” ready to run on any device.

The name echoes “liquid neural networks,” a family of neural networks at the heart of the company’s research.

Liquid Foundation Models

Liquid AI’s models are called LFMs (Liquid Foundation Models), with successive generations such as LFM2 and LFM2.5. They are hybrids: rather than relying only on the Transformer architecture, they mix convolution blocks and attention blocks. The aim is to use little memory and compute, so they can run on a phone, a laptop, a car or a connected device.

The documentation groups the models into six categories:

• text models (chat, tool calling, structured output);
• vision and audio models;
• decision models, for classifying or routing requests;
• Liquid Nanos, very small task-specific models (extraction, summarization, translation);
• encoder models, for search and embeddings.

Open, with conditions

LFM weights are published on Hugging Face, in formats suited to local inference (GGUF, MLX, ONNX). They are therefore open-weight models, but not open source in the strict sense: the LFM Open License v1.0 limits free commercial use to organizations below a revenue threshold. In practice, a startup or independent developer can integrate them freely, while a large company needs to negotiate a license.

How to access them

• Hugging Face, to download the weights, which work with llama.cpp, Ollama, MLX, vLLM or SGLang.
• Liquid’s Playground and OpenRouter, to test them through an API.
• The Apollo app on iOS and Android, for on-device AI.
• The official documentation, which also covers a managed fine-tuning service.

What sets Liquid AI apart is its “small and local” bet: less raw power, but AI that keeps your data on the device. To compare LFMs with other open models, see our page on open-source models.