Liquid AI is a US AI company founded in 2023 as a spinout of MIT CSAIL, commercializing liquid neural networks (LNNs), an AI architecture developed in the lab of Daniela Rus. The company markets LNN-based models as compact, causal, and energy-efficient alternatives to transformers, oriented toward on-device and edge deployment. Its flagship product line is the open-sourced Liquid Foundation Models (LFMs).
| Founded | 2023 (as an MIT spinout) | |
| HQ | Boston, MA | |
| Co-founders | [[daniela-rus | Daniela Rus]] (MIT CSAIL Director), Ramin Hasani, Mathias Lechner, Alexander Amini |
| Company type | Applied AI |
Snapshot
Funding
| Date | Amount | Round / Event | Source |
|---|---|---|---|
| 2024 | ~$250M Series A (reported) | Led by AMD Ventures | (public reporting; verify against primary filings) |
Funding figures are compiled from secondary sources and are treated as medium confidence until primary filings confirm them.
Technology
Liquid AI commercializes liquid neural networks (LNNs), an architecture developed in Rus's lab at MIT. The company characterizes LNN-based models along four properties:
- Compact — orders of magnitude smaller than transformers for comparable tasks.
- Causal — described as provably learning the task rather than the task context, supporting zero-shot generalization across environments.
- Energy efficient — optimized for on-device execution.
- Dynamics-based — computation inspired by the neural systems of small organisms such as C. elegans.
The architecture is positioned explicitly as not transformer-based, competing on generalization and efficiency rather than raw parameter count. The company places LNNs within the broader family of non-transformer architectures, adjacent to state-space models such as Mamba and LinOSS. In its own framing, LNN-based models compete with transformers on inference cost, claiming performance parity at a fraction of the compute.
Two research demonstrations are cited for the causal-learning claim. LNNs trained for drone navigation in summer woods are reported to transfer zero-shot to winter, fall, and urban environments without retraining, which the company presents as evidence of causal task-learning. Separately, on-device LFMs are described as able to describe photos and read text aloud locally, a capability relevant to accessibility applications for blind and visually-impaired users.
Products
The flagship product line is the Liquid Foundation Models (LFMs), open-sourced foundation models reported to be trained using roughly hundreds of GPUs, compared with tens of thousands for frontier transformers, and reported to perform on par on vision-language tasks. The company also pursues domain-specific LNN deployments in healthcare, aviation, and manufacturing, applications where edge deployment and explainability are emphasized. The LFMs are open-sourced as part of an ecosystem-building strategy.
Strategic position
Liquid AI is positioned as a principal commercial vehicle for the Private Physical Edge AI paradigm articulated in Rus's Digitalist Papers Vol. 2 essay (Private Physical Edge AI). The company's commercial argument rests on energy economics: in a setting where Eric Schmidt estimates the US needs roughly 92 additional nuclear plants' worth of energy for frontier AI, the company argues that compact, efficient models become strategically valuable. Its proposed market model is analogous to ARM's: licensing efficient AI cores to device manufacturers, with value shifting from hyperscalers toward applications and endpoints. The company describes itself as complementary to, rather than competing with, hyperscale frontier labs on frontier capabilities.
Relationships
- depends-on: Daniela Rus — co-founder, technical lead.
- supports: Edge AI / Private Physical AI — primary concept page.
- deploys-in: AI Robotics / AI Hardware Devices — application domains.
- related: Open-Source AI / Open-Weight Models — open-sources LFMs.
- related: AI Environmental Impact — energy-economics argument.
- contrasts-with: OpenAI / Anthropic / Google DeepMind — hyperscale frontier-lab paradigm.