An MIT CSAIL spinout building efficient, general-purpose AI that runs on phones, laptops, wearables and cars - fast, private, and off the cloud.
While the industry races to build ever-larger models in ever-larger data centers, Liquid AI is going the other direction - down, onto the hardware people already own.
Liquid AI is a foundation-model company with an unusual thesis: the most useful artificial intelligence will not necessarily be the biggest, and it will not always live in the cloud. Spun out of MIT's Computer Science and Artificial Intelligence Laboratory in 2023, the company builds what it calls Liquid Foundation Models - a family of general-purpose models designed to run efficiently and privately on-device, from smartphones and laptops to wearables, drones and cars.
The name is not marketing. It traces to a line of academic work called "liquid neural networks," an architecture the founders helped pioneer that draws on the nervous system of the roundworm C. elegans - an organism with roughly 302 neurons. The "liquid" refers to the model's ability to adapt its behavior after training, not just during it. In practice, that research heritage points Liquid toward models that are smaller, faster and cheaper to run than conventional transformer stacks of comparable capability.
The company describes itself, plainly, as "an efficiency-first foundation model company." Its mission is to build capable, general-purpose AI at every scale - models that are compute-optimized, aligned and, notably, explainable. That last quality is a deliberate contrast with the black-box reputation of frontier systems: one of Liquid's stated values is to "be white-box explainable," favoring transparent mechanisms over opaque shortcuts.
Latency, cost, and privacy are the three taxes of cloud-only inference. Liquid AI's pitch is to remove all three by moving the model to the metal.
Every prompt sent to a cloud model makes a round trip - out of the device, into a server, and back. That trip costs milliseconds you feel, dollars that compound at scale, and a copy of your data that leaves your control. For a chatbot on a laptop, that is tolerable. For a car making a driving decision, a factory sensor, or a pair of smart glasses, it often is not.
Liquid AI's answer is to make models small and efficient enough to run locally, so inference happens on the device with no network dependency. That flips the three taxes into features: near-zero latency, no per-token server bill, and data that never leaves the hardware - a genuinely private-by-design posture rather than a policy promise.
Where it differs from competitors is less about a single benchmark and more about design philosophy. Rather than shrinking a large transformer after the fact, Liquid built its architecture from first principles around efficiency, long context, and constrained hardware. And it pairs the models with deployment tooling - the hard, unglamorous part that decides whether on-device AI is a demo or a product.
It is not alone on the edge. Google's Gemini Nano, Microsoft's Phi, Meta's smaller Llama variants, Mistral's compact models and Qualcomm's on-device stack all compete for the same space. Liquid's wager is that a purpose-built architecture plus a frictionless deployment platform is a sharper combination than a scaled-down general model.
Liquid ships open-weight models, a deployment platform for developers, and a consumer app that proves the point - AI that runs entirely on your phone.
A family of general-purpose generative models built on a first-principles architecture rather than a standard transformer, tuned for strong quality at low memory and compute cost.
Since 2024Second-generation on-device models optimized for speed and memory, spanning base, instruct, vision-language and audio-language variants for real on-device agents.
2025 - 2026The Liquid Edge AI Platform and SDK deploys foundation models into iOS and Android apps in roughly 10 lines of code - no cloud infrastructure required.
2025A fully local, privacy-preserving AI chat app that runs LEAP's model library on-device. Works with no internet connection; nothing leaves the phone.
2025Open-weight models and a hosted playground for testing and evaluating LFMs, seeding a developer ecosystem around the edge stack.
Since 2024Efficient vision-language models built for the edge, extending Liquid's on-device reach from text to images and beyond.
2025Distribute open weights to win developers; monetize enterprise deployment, custom models and OEM licensing.
Liquid AI runs a two-sided playbook. Open-weight releases on Hugging Face and the LEAP platform build a developer community and put the models in front of the people who will embed them. Revenue comes from the commercial side: enterprise engagements, custom model work, and OEM licensing for companies that want efficient AI running on their own hardware - rather than a purely API-metered cloud business.
The customer list skews toward the physical and the regulated: automotive, consumer electronics, telecom, financial services and defense-adjacent sectors, alongside a broad base of mobile and edge developers. Reported engagements include Mercedes-Benz, Shopify, G42 and smart-glasses maker Brilliant Labs.
FIG. 1 - DISCLOSED FUNDING TO DATE. VALUATION ~$2.35B POST SERIES A (DEC 2024). BAR WIDTHS ARE ILLUSTRATIVE.
The founding team carried the liquid neural network thesis out of MIT CSAIL and into a company.
Persian-Austrian AI researcher and principal author of the Liquid Time-Constant Networks paper that put liquid neural networks on the map.
Helped originate the idea at TU Vienna's Radu Grosu lab before scaling it at MIT CSAIL.
MIT researcher and co-author on the liquid networks line of work spanning robotics and autonomy.
Robotics pioneer who leads one of the world's foremost AI labs and refined the liquid networks research at MIT.
A decade of research compressed into a fast-moving company.
Hasani, Lechner, Amini, Rus and colleagues publish the research that puts liquid neural networks on the map.
MIT News highlights liquid networks that keep learning and adapting after training.
The MIT CSAIL spinout launches with $37.5M in seed funding to build a new class of AI models.
The company raises at a ~$2.35B valuation to scale general-purpose, efficient AI.
Liquid ships second-generation on-device models, a deployment platform, and a fully local chat app.
A compact multimodal model family arrives for real on-device agents - text, vision and audio.
Hear the thesis in the founders' own words.
Official channels and further reading.