Intelligence on the Edge: Liquid AI's Ramin Hasani on the Search for Device-Native Foundation Models

Summary

• Liquid AI’s commercial thesis is that edge hardware is an underused inference market, not merely a cheaper copy of cloud AI. Hasani sized smartphones alone at roughly $500 billion annually; Labenz framed phones plus laptops as about $1 trillion of compute shipped each year, while the introduction used roughly $800 billion. Energy limits, privacy, latency, and workload economics all support Liquid’s aim to “build an intelligence layer on top of the diverse formats of hardware” already in pockets, cars, factories, and other devices.
• Liquid’s central technical finding is that the best architecture depends on scale, specialization, and the hardware constraint. Attention remains the richest, least-structured mechanism and may justify its (n^2) cost at frontier scale, while smaller or narrower models benefit from recurrence, convolutions, gating, or domain-specific dynamics. “The larger the network becomes, the more unstructured you can make it”; conversely, constrained systems can trade generality for substantially better speed and memory efficiency.
• The scientific lineage began with biologically inspired differential equations that produced striking control systems from only tens of neurons. Liquid networks parallel-parked a small car with 12 neurons, drove with 19, and flew a drone with 30; later work extended the approach to jets and other predictive systems. Their advantage was input-dependent dynamics and out-of-distribution adaptability—not magic or continual learning: “There’s no free lunch,” and the trained parameters remain fixed.
• Liquid’s automated foundation-model design system turns architecture selection into a hardware-grounded search problem. AFMD evaluates roughly 50–100 operators and hybrid combinations using an evolution strategy, actual target processors, memory and latency constraints, and around 100 downstream benchmarks rather than perplexity alone. After searches spanning roughly 10 million to 72 billion parameters, the lesson was “You have to give it to the algorithms,” including when the algorithms discard Liquid’s founders’ own preferred mechanisms.
• LFM2’s winning CPU design is surprisingly simple: 70–80% double-gated 1D convolutional layers, plus a smaller allocation to attention. The gate makes computation input-dependent; the convolution supplies a cheap, unstructured operator, replacing much of attention’s memory and quadratic cost without the elaborate hand-tuned machinery found in many alternative architectures. Hasani’s blunt account of the search result: “All of this has to go away” when extra gates and human-selected features fail the full efficiency objective.
• Commercial proof points suggest Liquid has moved beyond an architecture experiment. Hasani reported more than 1 million weekly Hugging Face downloads, the No. 5 position among U.S. organizations behind Google, Meta, Microsoft, and NVIDIA, more than 50 model instantiations used in enterprises, and only about 1,000 in-house GPUs. Shopify uses Liquid models in production across commerce workloads, Mercedes-Benz signed a contract for a roughly 600-megabyte in-car audio and visual system, and Labenz found the 1-billion-parameter Apollo model fast enough on an iPhone for private document search and classification.
• Silicon vendors may need to own a tunable intelligence layer, not stop at chips and kernels. Liquid is working with companies including AMD and Qualcomm around processor road maps, while Hasani points to NVIDIA’s Nemotron effort as evidence that models optimized for hardware can improve the enterprise proposition. Labenz challenged whether this creates excessive model-hardware coupling; Hasani’s answer was essentially, “Why do you want to change the model?”—provided the default is fast, tunable, and does not block the wider open-source ecosystem.
• The immediate local-agent opportunity is orchestration, while true intelligence-per-watt miniaturization needs new learning paradigms. A proposed local LFM2 24B A2B system can route requests, filter PII, invoke specialized models, and escalate difficult work to the cloud, but Hasani says no current local model matches frontier quality without fine-tuning; forthcoming tooling could create production models for tens to low thousands of dollars. Longer term, current architectures will not approach the brain’s roughly 20-watt efficiency because “intelligence for me is an emergent property,” requiring objectives that induce multiple ways of learning—not next-token prediction alone.

Deep dive

1. Liquid began by treating efficiency as the objective

• Hasani traces Liquid AI to roughly a decade of MIT research before the company spun out three and a half years earlier. The original objective was “maximizing the amount of intelligence that we can pack into the smallest format of algorithms,” with CPUs, NPUs, and small GPUs—not data-center clusters—as the assumed deployment environment.

• Robotics made efficiency inseparable from reliability. A car, drone, or fixed-wing vehicle quickly encounters distribution shift, so the research question was whether compact systems could carry learned concepts into conditions they had never seen rather than merely interpolate within training data.

• That lineage became a broader mission: build efficient, general-purpose systems “at every scale,” extending from physical control and predictive sequences into language, audio, vision, and video. Efficiency remained the organizing constraint even as Liquid moved from specialized neural circuits to billion-parameter foundation models.

2. A 300-cell worm inspired adaptable control with 12 neurons

• In 2015, Hasani and co-founder Mathias Lechner studied C. elegans because it was the only animal whose complete nervous system they knew, yet its roughly 300 cells generated extensive sensory-reactive control. The disparity with much larger robotic networks suggested biology was using a more expressive computational unit.

• The worm’s neurons are graded rather than spiking, making their behavior differentiable enough to connect with artificial neural-network training. Liquid’s researchers assembled circuits of two, four, eight, and eventually hundreds of differential-equation-based neurons, then optimized them with backpropagation.

• Each node had richer internal feedback and nested nonlinearities than a conventional artificial neuron. “Liquid” referred to adaptability: even after training, the system’s dynamics remained flexible and input-dependent, creating more degrees of freedom for responding to unfamiliar conditions.

• The demonstrations remain the sharpest proof of compression: 12 neurons parallel-parked a small car, 19 drove one, and 30 navigated a drone. Work involving the U.S. Air Force and a Boeing-supported MIT postdoc also showed that “a handful of these neurons” could control jets under changing conditions.

3. Liquid dynamics add capacity without continual learning

• Labenz noticed that Hasani counted neurons where foundation-model builders count parameters. For the early liquid formulation, Hasani offered a rough conversion: multiply neurons by seven to estimate parameters, reflecting the multiple gates and feedback paths inside each cell. Modern LFMs use conventional parameter counts; a 1-billion-parameter LFM means the same thing as elsewhere.

• The extra capacity comes from dynamics as a second axis beyond parameter count. Inputs alter how the fixed system evolves, allowing a small network to encode more adaptive behavior, but its trained weights do not update during deployment. Hasani therefore distinguishes liquid adaptation from continual learning, which “continuously receives new data and retunes itself.”

• His concrete example was rain striking an autonomous car’s camera. The scene’s underlying variables have not changed, but the image contains unfamiliar noise; liquid dynamics can react differently and apply low-pass filtering. That responsiveness may improve robustness, yet it does not mean the network learns new parameters while driving.

4. Closed form removed the solver but not sequential computation

• The original equations were expensive because nonlinear recurrence is difficult to turn from sequential scalar or vector operations into parallel tensor computation. Linear state-space models scale precisely because they can parallelize the dynamics, then apply a pointwise nonlinearity; nested nonlinear relationships between parameters resist that conversion.

• Hasani situated the problem in a lineage beginning with Louis Lapicque’s 1907 membrane-potential equation and Hodgkin and Huxley’s later biological model, begun in 1953 and recognized with a 1963 Nobel Prize. Textbooks treated the relevant equation family as having no known closed-form solution.

• Liquid’s team reported such a solution around 2022, publishing its closed-form continuous-time work in Nature Machine Intelligence that November. Removing numerical solvers meant the equations no longer required repeated approximation steps and could, mathematically, be assembled at far greater scale while retaining their nonlinear character.

• Practical parallelism remained the ceiling. Hasani estimated closed-form systems could move from hundreds to hundreds of thousands, perhaps 1–10 million neurons, but still run sequentially because of nested nonlinearities. Many fit in roughly 1–25 megabytes and run on a CPU or Raspberry Pi; ongoing scan research seeks subquadratic speedups without linearizing them.

5. Scale determines how much architectural bias survives

• Transformers win at maximum scale because attention and matrix multiplication are unusually unstructured: nearly any matrices can be multiplied in parallel without imposing strong assumptions about the task. Hasani suggested that at trillions or tens of trillions of parameters, attention’s (n^2) cost “maybe” is genuinely necessary for the richest capabilities.

• Smaller models occupy a different regime. Recurrence, gates, convolutions, and specialized nonlinearities impose biases that can improve learning efficiency when data, parameters, latency, or use cases are bounded. Hasani put the potentially useful range for alternatives as high as 100 billion or perhaps 1 trillion parameters, while treating those boundaries as regime-dependent rather than universal.

• Labenz’s “bitter-pilled” question was why the original liquid paradigm could not simply scale to billions of neurons. Hasani’s answer was that structured nonlinearity blocks efficient parallelization; at frontier scale, “as soon as you start adding a little bit of bias,” the structure can become a liability.

• The resulting gradient is conditional rather than doctrinal. If raw speed matters enough to sacrifice some quality, a fully linear system may win; if maximum intelligence is the goal, use the least-biased architecture available. Humanity’s resource constraint is what makes the middle ground—attention hybrids and specialized operators—commercially important.

6. Small models reached large distribution with modest compute

• Hasani reported that Liquid had reached more than 1 million Hugging Face downloads per week and ranked fifth in U.S. downloads, behind Google, Meta, Microsoft, and NVIDIA. He contrasted that distribution with roughly 1,000 GPUs in house and more than 50 model instantiations that people were using in enterprises.

• Labenz’s own test supplied a useful quality boundary. Liquid’s Apollo app runs a 1-billion-parameter hybrid model fast enough on an iPhone to search and classify local documents privately, making it practical for basic workflows, though he explicitly described it as “far from the frontier.”

• The signal is therefore not that a 1-billion-parameter device model replaces the largest cloud systems. It is that a sufficiently fast, private model can become the economically rational option for filtering, retrieval, classification, and other bounded work that does not require “the fanciest type of intelligence.”

7. AFMD replaces architectural taste with hardware-grounded search

• Liquid’s Automated Foundation Model Design system, or AFMD, is a meta-learning process using an evolution strategy with target hardware in the loop. Its objective balances memory consumption, latency, and speed while requiring no sacrifice on the downstream quality criteria chosen for the deployment.

• Perplexity was not reliable enough. Liquid evaluates candidates on the actual applications they must perform, across roughly 100 benchmarks, and on the processor where they will run. A theoretically attractive operator can lose once memory movement, kernel support, latency, and customer-specific quality enter the same objective.

• Hasani calls influential researchers inside foundation labs the “Avengers of the architectures”: small groups make consequential tweaks because something worked in their personal experiments. Liquid built AFMD to remove that cultural bottleneck and its own founders’ attachments alike. “You have to give it to the algorithms. You have to have Bitter Lesson people.”

• The search included roughly 50–100 attention, convolutional, recurrent, state-space, and liquid-derived operators, assembled into hybrids. Liquid ran scaling studies from about 10 million to 72 billion parameters, repeatedly asking which computational graph delivered the most accuracy per unit of memory and latency on a particular processor.

8. LFM2 found that simple gated convolutions can displace most attention

• For CPU deployment, AFMD converged on a simpler graph than the founders expected. Roughly 70–80% of LFM2 consists of unstructured 1D double-gated convolutions, with a small number of attention layers retained.

• The host’s intuition for gating was essentially correct: learn a transformation, then let a smaller mechanism modify it for the current input. Hasani called that “a liquid structure,” because the transformation is no longer fixed across examples; the system has learned how its computation should vary with incoming data.

• The important effect appears during learning, not merely the forward pass. Gradients flowing through an input-dependent operator teach a representation of dynamics as well as static parameters. Hasani connected this to Liquid-S4, which he said introduced input-dependent state-space models roughly a year and a half before Mamba.

• The search rejected many hand-tuned additions used in Mamba-style systems, gated delta networks, and linear-attention variants. “All of this has to go away” was Hasani’s summary of what is required for the most efficient architecture, leaving the double-gated convolution as the key surviving addition. Whether even that gate remains useful at 100-trillion-parameter scale is, he stressed, an open question.

9. Architecture optimizes efficiency, not intelligence by itself

• Hasani’s strongest claim for architectural research is narrower than “architecture creates AGI”: it makes computation efficient without losing quality. Future intelligence also depends on data representation, memory, priors, objectives, and learning algorithms—especially how a model receives signals across long histories rather than only predicting the next token.

• He described current systems as largely i.i.d. “averaging machines,” with recognizable effects on writing and sequence generation. Multi-agent systems and test-time scaling may compensate for some limitations, but he argued that the “definition of learning theory itself is broken at scale” and needs deeper revision.

• Labenz proposed that new learning paradigms will lead and architecture will subsequently make them efficient. Hasani’s correction was worth preserving: architecture is “one component,” not merely an implementation afterthought. Recursive self-improvement, meanwhile, is a newer framing for continual-learning research that has existed for decades.

10. Narrow data rewards architectures shaped like the problem

• Biological sequences illustrate why attention is not universal. Their vocabulary is small while relevant context may span “one to 100 billion” elements; at those lengths, quadratic attention becomes prohibitive. Pure convolutions, state-space models, liquid networks, parallel recurrence, or linear attention can exploit the simpler alphabet and long-range structure.

• For audio-only signal transformation—voice to voice, noise to signal—recurrent systems remain powerful, especially with limited data. Architectural feedback supplies a useful prior: “the closer the architecture is to the dynamics of the data set,” the less the learner must infer from examples alone.

• Video may require a different learning prior. Hasani pointed to diffusion’s success while acknowledging continuing disagreement over whether autoregressive modeling can ultimately achieve the same result. He treated diffusion as straddling the boundary between architecture and learning algorithm rather than forcing a clean taxonomy.

• Extreme latency and physical simulation push specialization further. Microsecond control cannot afford a large generic graph, while a factory’s chemical-reaction digital twin may naturally favor physics-informed or continuous-time differential-equation models. Original liquid networks remain useful for sensor, physical, and multivariate predictive sequences precisely because their dynamics match the data.

11. Device AI converts idle silicon into a second inference market

• Liquid defines “device foundation models” as intelligence for processors outside data centers; inside data centers, it targets constrained workloads such as ultra-low latency, very long sequences, small memory footprints, and cost-sensitive inference. Each open LFM generation is meant to beat the prior one on the criteria Liquid cares about without sacrificing quality.

• Shopify already runs Liquid models in production across areas including recommendations, search, product-catalog understanding, and multimodal commerce. Hasani said the models improve click-through rates and other internal measures, grounding the architecture story in customer outcomes rather than public benchmarks alone.

• A Mercedes-Benz contract takes the thesis outside the data center: Liquid models are intended to power audio and visual intelligence inside the car. Hasani said the roughly 600-megabyte system can fit on a small in-car processor while offering voice quality comparable to much larger audio systems.

• The market figures varied by framing but pointed in one direction. Hasani put smartphones at roughly $500 billion annually; Labenz described phones plus laptops as about $1 trillion of yearly compute, calling it “a lot of dark compute.” Local inference also conserves scarce energy by reserving frontier cloud models for problems that genuinely require them.

12. Hardware vendors need an intelligence layer above their kernels

• Heterogeneous devices make optimization a multilayer problem. Open GPU kernels provide material for automated kernel engineering, but proprietary NPUs expose less of their design and may require a separate hardware-grounded search. Matrix multiplication alone can have scores of implementations depending on caching, memory, and CPU/GPU workload division.

• Liquid searches architecture before pre-training, then applies post hoc work such as quantization-aware training, bit-width changes, kernel optimization, and processor-specific scheduling. Because a full model run can cost millions of dollars, the goal is to anticipate inference behavior before committing that training budget.

• Work with AMD, Qualcomm, and other semiconductor companies extends into road maps and possible next-generation ASIC choices. Hasani argues that kernel optimization is becoming automatable; hardware vendors should move upward to an “intelligence layer” naturally fitted to their silicon, citing NVIDIA’s Nemotron work as the strongest business example.

• Labenz asked whether tight optimization means a future of vertical integration where models become hard to swap. Hasani’s answer—“Why do you want to change the model?”—defended a fast, tunable default, while still saying vendors should optimize for open-source models. He also thinks glasses “could” become a major computing medium, perhaps eventually competing with laptops.

13. Local agents will route private work before they rival the frontier

• Liquid’s local co-work demonstration uses LFM2 24B A2B—a mixture-of-experts model with 2 billion active parameters out of 24 billion total—to show what consumer hardware can host. Hasani was explicit that the off-the-shelf model does not yet perform every requested workflow at frontier quality: “None of the local models today are there.”

• His envisioned local computer is an orchestrator rather than one omnipotent model. It routes hard, nonsensitive tasks to cloud systems; invokes small local models for bounded work; filters personally identifiable information before transmission; and coordinates tools and services. “That router is the computer.”

• For Labenz’s five-year personal database, the model worth tuning is principally the orchestrator: it must know which local search, privacy filter, specialized model, or cloud model to call—and recognize its own shortcomings. Today that requires fine-tuning and production checks; a frontier cloud agent cannot simply manufacture the finished local model without access to an automated tuning platform.

• Liquid is building such a platform and hoped to announce it “in the next few months.” Hasani estimated production-grade customization at tens of dollars to low thousands, not tens of thousands. A terminal call could initiate fine-tuning, potentially using synthetic data so the service need not inspect the user’s private corpus directly.

14. Human-level intelligence per watt needs new ways to learn

• Hasani does not believe transformers or today’s alternative architectures can approach the human brain’s intelligence at roughly 20 watts. The comparison must also account for biological evolution: humans do not read the entire internet before reasoning, but their inherited machinery reflects an extraordinarily long and energy-intensive design process.

• Next-token prediction unexpectedly produced in-context learning—an “emergent property” Hasani characterizes as a vague least-squares algorithm or “gradient descent in a mushy way.” Humans appear to have many such internal procedures: learning from examples, reinforcement, mental simulation, Bayesian reasoning, and other mechanisms triggered as circumstances require.

• He doubts that explicitly forcing trajectory-based reinforcement learning into current systems is sufficient. The deeper research problem is to discover foundational objectives—next-token prediction being only one—that cause reinforcement, curiosity, and other learning algorithms to emerge while using limited energy in the final deployed system.

• Hasani closed as a scientist still training experiments while serving as CEO. Agents now let more people attempt frontier research, but realizing that opportunity requires “rebasing our biases” about work, automation, and armies of agents. His positive vision is a society organized more around science’s original purpose: understanding the world by satisfying curiosity.