⚡️ A Worm With 302 Neurons Inspired Their Architecture — Ramin Hasani, Liquid AI

Ramin Hasani explains that Liquid AI’s research began roughly 12 years ago at the Vienna University of Technology, where he and co-founder Matthias pursued a first-principles way to bring continuous-time mathematics into machine learning. They studied how neurons exchange information and modeled those dynamics with first-order differential equations. Human brains were too poorly understood and difficult to learn from, so they turned to *C. elegans*, a two-millimeter worm with just 302 nerve cells and 95 muscle cells. The worm was useful because its neurons do not spike like conventional biological neurons; they have a graded, continuous behavior that resembles artificial neurons. Hasani says the worm could control its muscles better than contemporary robotic systems despite its tiny nervous system. He also notes that humans and the worm split roughly 600 million years ago, that the worm’s genome is about 78% similar to humans’, and that research on it has contributed to four Nobel Prizes. Modeling its neural dynamics produced liquid time-constant neural networks: flexible, nonlinear, differentiable systems that can be trained with backpropagation. Their small size also made them attractive for robots, where tens or hundreds of neurons could provide useful autonomous control on limited onboard compute.

In 2017, Daniela Rus of MIT’s Computer Science and Artificial Intelligence Laboratory contacted Hasani’s professor and invited Hasani and Matthias to continue their work at MIT in real vehicles, including cars and drones. At MIT, Hasani met co-founder Alexander Amini, and the group worked on continuous-time dynamical systems, sequence modeling, and applying these methods to robotics. Their research focused on systems that could generalize outside the distribution of their training data while remaining small and computationally efficient. Hasani describes continuous-time models as an extension of recurrent networks: they account not only for the sequence of events but also for the time between events. Around 2020, however, scaling these nonlinear recurrent systems became a major problem. Transformers scale well because attention can be expressed through highly parallel matrix multiplications, although its cost grows quadratically with more context. Recurrence can offer subquadratic or near-linear computation, but nonlinear recurrent operations are difficult to parallelize on GPUs. State-space models address that hardware problem by linearizing the dynamics, enabling parallel scans and linear-algebra techniques. The tradeoff is expressivity: Hasani says linearized systems can struggle with long-context dependencies and generally fall short of transformers on language tasks.

Rather than claiming that liquid networks were a universal replacement for transformers, Liquid AI assembled researchers associated with several alternative architectures, including Hyena, state-space models, and liquid neural networks. The company built a meta-AI system that searches through combinations of operators and selects a hybrid architecture for a target deployment environment, such as a GPU or an NPU. The search optimizes four criteria: preserving quality relative to a pure transformer, reducing memory consumption, reducing latency, and increasing computation speed. Hasani describes the resulting framework as a systematic, recursively improving architecture search and points to a paper called STAR—“Systematic Design of Tailored Hybrid Architectures”—as the scientific account of the approach. The first Liquid Foundation Model was a mixture of convolutions, attention, and liquid components. The LFM 2 generation was optimized particularly for CPUs and uses a simpler structure: about 80% double-gated one-dimensional convolutions and 20% grouped-query attention. Liquid has scaled its model family from tens of millions to 70 billion parameters, while releasing smaller LFM 2 and LFM 2.5 models and preparing LFM 3 with improved quality, stability, and latency control.

Hasani rejects the idea that Liquid is only a research project. He says Shopify tested the models directly and was positively surprised by their quality and efficiency. Liquid models are deployed across Shopify’s merchant- and customer-facing products; one example is the Shop app, where searches send signals to a Liquid foundation model. Shopify’s Liquid deployments now process more than one billion requests per month. The open models have received more than 40 million downloads overall, with about 1.5 million downloads per week. Liquid is also targeting intelligence outside data centers, working with AMD on AI PCs. Hasani estimates that roughly 300–400 million laptops are produced annually, with up to 100 million potentially AMD-powered. Local models could provide privacy filtering, decide what sensitive information may be sent to cloud systems, and act as offline assistants with access to information on the user’s computer. He says on-device AI is driven partly by cost: sending requests from hundreds of millions or billions of devices to the cloud is expensive. But achieving production quality requires more than a benchmark; teams must handle thermal behavior, long-context changes, reliability, and the realities of deployment. Mercedes-Benz is another example. Liquid’s roughly 600-megabyte multimodal model is planned for deployment in North American Mercedes vehicles of the relevant generation during the year discussed. It can support voice interaction with the car, connect to cameras, and combine audio, vision, and text-based reasoning on a chip costing about $100.

Liquid’s models are also being used in robotics experiments. Hasani cites RoboTech and an Italian company, G-Bionics, whose modular robots use Liquid’s vision-language models for instruction following and vision-to-action behavior. Cameras on the robots provide visual input, and the models help determine actions, especially in industrial and factory settings. Smaller multimodal models are particularly useful where low latency matters. Despite Liquid’s origins in robotics, Hasani says the company deliberately postponed making robotics its primary go-to-market vertical. Production robotics requires extensive verification and validation, followed by regulatory work, which makes sales cycles and deployment slower. Liquid instead built a horizontal intelligence layer for OEMs and other customers first, hoping to enter more sensitive action-oriented applications after its models, data controls, post-training, and customization processes matured. Asked whether robotics needs fundamentally different models, Hasani separates three kinds of innovation: model architecture, learning algorithms, and data processing. He says an LFM architecture could be used in a world-modeling pipeline and trained to learn physical structure. World modeling extracts different primitives from the physical world than next-token prediction does, so the same general-purpose architecture may be useful, while the training objective and data pipeline change. He considers the question of which learning algorithms will produce the “ChatGPT moment” for robots unresolved.

Liquid typically begins an enterprise engagement with a design partner. Its applied machine-learning team works with the customer and charges for that engagement; once the solution is deployed, Liquid charges recurring annual, per-device licensing fees. Hasani says a first customer such as Mercedes helps Liquid develop a vertical solution, after which later automotive customers require less service and can be deployed more quickly. The company is building a beta platform that automates more of this model-development process. It is intended to let enterprises perform pre-training, mid-training, post-training, data generation, reinforcement learning, and other forms of customization rather than merely fine-tune an existing model. Hasani frames this as a shift from an industry focused mainly on inference tokens—hosting models and charging for their use—to a market for customization tokens, in which companies pay to build and continually adapt their own intelligence. The existing developer platform, Leap, is more modular. Developers can use Liquid’s cookbook, fine-tune models, and export inference-ready bundles such as GGUF packages suitable for deployment through tools like llama.cpp. Hasani says the new beta platform is an automated version of those capabilities, designed to work through an agent or developer workflow. The first deployment may be automated from evaluations, but he argues that production systems need continual adaptation because real-world data distributions change; static evaluations eventually become obsolete.

Hasani says Liquid’s architecture search runs scaling-law experiments for each candidate: the team increases data or model size and measures how well the candidate trains. One result is that larger general-purpose models should be relatively unbiased and simple. LFM 2, for example, does not include every specialized mechanism the team has studied; it relies heavily on straightforward one-dimensional convolutions. At smaller scales, however, adding structural biases and feedback loops can make a model more expressive when memory and compute are constrained. The best architecture also depends on the modality. Hasani says continuous-time systems and state-space models are particularly effective for audio and other sequential time-series data, but state-space models have been weaker on text. He believes current GPU hardware favors parallel matrix multiplication and therefore constrains which architectures are practical. Photonic computing could eventually change that tradeoff by making fast sequential or nonlinear computation more feasible. Quantum computing, in his view, would most clearly transform search and optimization by enabling fundamentally different ways to explore parameter spaces, though he recalls being told in 2017 that broadly usable quantum systems might arrive around 2035. Looking ahead, Liquid is researching multimodal systems that process audio, vision, and text together; longer-horizon reasoning; efficient kernels and infrastructure; biological data such as DNA; and models that can learn continuously. Hasani’s broader target is embodied, adaptive intelligence that can operate outside data centers and keep improving in its environment.