Liquid AI
Last reviewed
Sources
25 citations
Review status
Source-backed
Revision
v4 ยท 6,216 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Sources
25 citations
Review status
Source-backed
Revision
v4 ยท 6,216 words
Add missing citations, update stale details, or suggest a clearer explanation.
Liquid AI is an artificial intelligence company founded in 2023 by researchers from MIT CSAIL and headquartered at 314 Main Street in Cambridge, Massachusetts.[3] The company develops Liquid Foundation Models (LFMs), a class of large neural networks built on a non-transformer architecture inspired by Liquid Time-Constant (LTC) networks.[4][1] LFMs are designed to be more memory-efficient than transformers and competitive with them at lower compute and memory cost, particularly for long-context inference and on-device deployment.[15]
The company was spun out of the Massachusetts Institute of Technology Computer Science and Artificial Intelligence Laboratory by Ramin Hasani, Mathias Lechner, Alexander Amini, and Daniela Rus.[3] Its public profile rose sharply with the September 30, 2024 launch of the first three LFMs (LFM-1B, LFM-3B, and LFM-40B)[14][15] and the December 2024 close of a $250 million Series A round led by AMD Ventures at a reported valuation of about $2.3 billion.[6][12] By 2026 the company had announced production deployments with Mercedes-Benz for in-vehicle voice intelligence[11] and a multi-year commerce partnership with Shopify,[8] and its LFM2 family had become one of the most downloaded open-weight model families optimised specifically for phones, laptops, and embedded systems.[5]
Liquid AI sits at the intersection of three trends in foundation models: the search for transformer alternatives that scale linearly with sequence length, the push to bring capable models onto end-user devices instead of the cloud, and the resurgence of biologically inspired neural network research. The company's distinguishing claim is that small, dynamically evolving recurrent units patterned on the neurons of the nematode Caenorhabditis elegans can deliver transformer-class quality with a fraction of the parameter count and memory budget, and that this advantage compounds at the edge where battery, latency, and privacy constraints are tightest.[13]
Liquid AI grew directly out of academic work on continuous-time recurrent neural networks carried out by Daniela Rus's group at MIT CSAIL together with collaborators at the Institute of Science and Technology Austria (ISTA) and the Vienna University of Technology (TU Wien). The most important precursor paper, "Liquid Time-constant Networks" by Hasani, Lechner, Amini, Rus, and Radu Grosu, was presented at AAAI 2021 and posted to arXiv as 2006.04439 in June 2020.[1] A follow-up paper, "Closed-form Continuous-time Neural Networks," appeared in Nature Machine Intelligence in November 2022 and showed that the recurrent dynamics of LTC networks could be approximated in closed form, removing the need for an ordinary differential equation solver and accelerating training and inference by one to five orders of magnitude.[2]
The LTC architecture was partly inspired by the nervous system of Caenorhabditis elegans, a one-millimetre roundworm whose 302 neurons and roughly 7,000 synapses have been completely mapped.[13] Several of Hasani and Lechner's earlier papers used C. elegans-style circuits to control simulated cars and lane-keeping policies, demonstrating that very small continuous-time networks could perform tasks that normally required much larger feed-forward or recurrent networks. In one widely cited result, a 19-neuron liquid network learned to keep a simulated car in its lane using only sparse sensory input, a task that conventional deep networks typically need on the order of 100,000 artificial neurons and half a million parameters to perform.[17] That line of research provided the conceptual underpinning for the company's later claim that biologically inspired, dynamically evolving neurons can rival transformer-based models at much lower parameter count.
Hasani and Lechner incorporated Liquid AI in March 2023 with Amini and Rus as co-founders, and the company emerged from stealth later that year.[3] An initial seed round of about $37.6 million closed in December 2023 at a reported valuation of approximately $303 million, with foundation model research and enterprise consulting partner Capgemini among the investors alongside Automattic, Samsung Next, Bold Capital Partners, ISAI Cap Venture, OSS Capital, and PagsGroup.[21] The seed funding allowed the team to build a small but well-credentialed engineering and research staff in Cambridge and to begin training larger Liquid Neural Network variants under the LFM brand.
The choice of Cambridge, Massachusetts as headquarters was deliberate. Liquid AI's offices at 314 Main Street sit inside the Kendall Square cluster that surrounds MIT, putting the company within walking distance of the CSAIL building where most of its founding research was carried out.[3] The geographic proximity made it easier to retain MIT research affiliations for Hasani and Rus and to recruit from the same talent pool that supplies firms such as Boston Dynamics and DataRobot.
The four academic co-founders all spent significant time in Daniela Rus's MIT CSAIL group, although three of them did their PhD work elsewhere.[3] The team has remained largely intact since founding, with Hasani as CEO and Lechner as Chief Technology Officer.
| Name | Role at Liquid AI | Background |
|---|---|---|
| Ramin Hasani | Co-founder and CEO | PhD, TU Wien, 2020; postdoctoral researcher with Daniela Rus at MIT CSAIL; first author on the LTC and CfC papers;[1][2] previously principal AI scientist at Vanguard |
| Mathias Lechner | Co-founder and CTO | PhD, IST Austria, 2022, supervised by Tom Henzinger; MSc and BSc, TU Wien; research affiliate at MIT CSAIL |
| Alexander Amini | Co-founder and Chief Scientific Officer | PhD MIT CSAIL, 2022, supervised by Daniela Rus; lead lecturer of MIT 6.S191 Introduction to Deep Learning |
| Daniela L. Rus | Co-founder and board director | Director of MIT CSAIL since 2012; Andrew and Erna Viterbi Professor of EECS; PhD Cornell, 1992; MacArthur Fellow (2002) |
| Mathew Hein | Strategic AMD relationship lead (external) | AMD SVP Strategy, joined as AMD Ventures lead investor in Series A[6] |
Hasani's career trajectory is unusual for a CEO of a frontier AI company. He spent the early years of his postdoctoral work in Vienna and at MIT publishing on small continuous-time networks for control problems before stepping into a principal AI scientist role at the Vanguard Group, where he applied liquid networks to financial time-series analysis. He returned to academic affiliation at CSAIL in 2022, just as the closed-form continuous-time paper was being prepared for Nature Machine Intelligence,[2] and used that window to begin the company-building work that became Liquid AI. Lechner, whose dissertation at IST Austria focused on the formal verification of neural network controllers, brings the engineering rigour for which Tom Henzinger's group is known. Amini, the youngest of the four, is best known outside the company as the lead instructor for MIT 6.S191, the deep-learning class whose YouTube lectures have several million views.
Mikhail Parakhin, the former head of Microsoft Bing, Windows, and Ads and currently CTO of Shopify, has publicly endorsed Liquid AI as the front-runner among non-transformer foundation model startups and is widely associated with the company in an advisory capacity, but he has not become CEO; Hasani remained CEO through 2026. Reports that Parakhin took the CEO role appear to confuse his Shopify CTO position with a leadership move at Liquid AI.
Liquid AI's funding history has been short and concentrated. Two rounds account for nearly all disclosed capital raised, and AMD's strategic involvement in the Series A doubles as an engineering and chip partnership rather than a pure financial bet.[6]
| Round | Date | Amount | Lead investor | Selected participants | Reported valuation |
|---|---|---|---|---|---|
| Seed | December 2023 | $37.6 million | OSS Capital and PagsGroup (co-led) | Capgemini, Samsung Next, Automattic, Bold Capital Partners, ISAI Cap Venture, Tom Preston-Werner, Tobias Lutke, Bob Young | Approximately $303 million[21] |
| Series A | December 13, 2024 | $250 million | AMD Ventures | Duke Capital Partners, OSS Capital, PagsGroup, Tom Preston-Werner, Shopify, others | Approximately $2.3 billion[6][12] |
The seed cap table is worth a closer look because it foreshadowed many of the company's later commercial relationships. Capgemini, the French IT services group, became the launch partner for enterprise deployments of Liquid Neural Networks in regulated industries.[9] Samsung Next's involvement preceded the Hyena Edge demonstration on the Samsung Galaxy S24 Ultra by about 18 months.[22] Tobias Lutke, the Shopify CEO, was an angel investor at the seed stage, and Shopify itself returned at Series A and signed a multi-year commercial agreement in late 2025.[8] Tom Preston-Werner, the GitHub co-founder, has been a vocal public proponent of LFMs on social media. Bob Young, the Red Hat co-founder, supplied open-source commercialisation experience that informed the later Liquid Open License.
The Series A is unusual for being led by the venture arm of a chip company rather than a generalist VC.[12] AMD Ventures committed not only to fund the round but to co-train and co-deploy LFMs on AMD Instinct GPUs and AMD Ryzen client processors, which became the basis of the LEAP edge platform announced in 2025.[6][10] According to the company's blog post announcing the round, the funds were earmarked for compute infrastructure, on-premises and edge inference, expansion into consumer electronics, telecommunications, financial services, e-commerce, and biotechnology, and continued research on LFM architectures across model sizes and modalities.[6] The roughly 7.6x markup between the seed and Series A valuations was a noticeable outlier in a December 2024 fundraising market that had already tightened around pure-play foundation model startups.
Liquid Time-Constant (LTC) networks are continuous-time recurrent neural networks in which each neuron's time constant depends on its input.[1] In a standard recurrent network, the hidden state evolves according to a fixed update rule. In an LTC, the hidden state x(t) evolves under an ordinary differential equation roughly of the form
dx(t)/dt = -[1/tau + f(x(t), I(t), t, theta)] * x(t) + f(x(t), I(t), t, theta) * A
where I(t) is the input, tau is a base time constant, f is a learned nonlinearity, A is a learned bias vector, and theta are the trainable parameters. The crucial difference from earlier neural ODE work, including the 2018 Chen et al. paper that introduced Neural Ordinary Differential Equations, is that the effective time constant 1/tau + f varies with both the hidden state and the input.[1] The network's response speed is therefore "liquid": slow when input is mild and fast when input is sharp, in a way that approximates the behaviour of biological neurons whose membrane time constants depend on incoming synaptic activity.[17]
LTCs are universal approximators for arbitrary continuous-time dynamics and can be trained by backpropagation through ODE solvers, although the original implementation depended on numerical integration that was expensive at training time.[1] The 2022 Closed-form Continuous-time (CfC) paper showed that a particular family of LTCs admits a closed-form solution, which means the same dynamics can be evaluated by a normal forward pass without an ODE solver.[2] This made the architecture practical for the kind of large-scale training that the LFM line later required.
In benchmark studies on irregularly sampled time series, gesture recognition, autonomous lane-keeping, and human activity recognition, very small LTC and CfC networks have repeatedly matched or beaten LSTMs and GRUs with one to two orders of magnitude more parameters.[2] That parameter efficiency is the main empirical claim that Liquid AI is now trying to scale up to natural language, vision, and speech.
Another property of LTC networks that motivates the company's edge focus is interpretability. Because each neuron is governed by a low-dimensional ODE rather than a high-dimensional dense projection, the dynamics of a trained LTC can be visualised and, in some cases, written out symbolically. Researchers at MIT have used this property to extract the causal structure that liquid networks learn from driving data, showing that a 19-neuron liquid controller attends to a small, interpretable set of road features rather than to the noisy periphery of the input.[17] That contrast with the black-box reputation of attention-based models is a recurring talking point in Liquid AI's external communication, although the property is not always preserved at the multi-billion-parameter scale at which the company now trains its language models.
The first generation of Liquid Foundation Models was released on September 30, 2024.[14][15] The launch covered three models intended to span on-device, edge, and cloud workloads.
| Model | Parameters | Architecture | Position |
|---|---|---|---|
| LFM-1B | 1.3 billion | Dense, hybrid Liquid layers | Resource-constrained edge inference |
| LFM-3B | 3.1 billion | Dense, hybrid Liquid layers | Mobile, robotics, drones |
| LFM-40B | 40.3 billion total, 12 billion active | Mixture-of-Experts | Cloud and on-premises servers |
The models are not pure transformers. They are built from "liquid" computational units inspired by the LTC and CfC work and combined with other operators including long convolutions in the Hyena family, gated convolutions, and a small number of attention layers.[15] The mix of operators differs by model size and is chosen so that long-context inference scales linearly with sequence length rather than quadratically as in a standard self-attention transformer. Liquid AI claimed at launch that the LFMs could handle context windows up to 32,000 tokens out of the box, with extension experiments suggesting close to one million tokens at modest memory cost.[15]
On benchmarks reported at launch, LFM-1B was claimed to outperform comparable 1 to 1.3 billion parameter transformer models on MMLU, ARC-C, and similar reasoning suites.[14] LFM-3B was positioned against Microsoft's Phi-3.5, Meta's Llama 2 13B and Llama 3.2 3B, and Mistral 7B, with the company saying the model was competitive on common-sense reasoning and instruction following while using a substantially smaller memory footprint at long context.[15] LFM-40B, the Mixture-of-Experts member, was pitched as roughly comparable to Llama 3.1 70B on standard tasks while activating fewer parameters per token, and posted a five-shot MMLU score of approximately 78.8 in the company's launch numbers.[14] Independent benchmarks at the time generally confirmed that the models were strong for their parameter count, though they did not match the very largest open-weight transformers across every task.
The initial release was English-focused. The models were available through the Liquid Playground hosted by the company and through partners such as Lambda Labs and Perplexity Labs.[15] Weights for the first generation of LFMs were not released openly, which became one of the more pointed criticisms of the launch. The opacity of the initial release also made it difficult for outside researchers to verify some of Liquid AI's more aggressive efficiency claims, particularly around context-length scaling beyond the official 32k window. That gap between marketing claims and reproducible artefacts narrowed considerably with the LFM2 line, which was released under an open licence and accompanied by a full technical report.[19]
In December 2024 Liquid AI introduced STAR, short for Synthesis of Tailored Architectures, a framework for automatically designing Liquid models.[16][24] STAR uses an evolutionary search over a hierarchical encoding the company calls "STAR genomes," which can describe combinations of attention heads, recurrent units, gated convolutions, Hyena-style operators, and Liquid blocks.[24] The search is guided by a multi-objective fitness function that trades off perplexity against latency, memory, and KV-cache size on a target hardware platform. According to the company's research blog, STAR was able to find architectures that reduced cache size by up to 37% relative to existing hybrid models and up to 90% relative to standard transformers while maintaining or improving quality on language modelling tasks.[16]
Under the hood, STAR represents a candidate model as a numeric sequence whose digits encode both the operator placed at each layer and the connectivity between layers.[24] The search space is grounded in the theory of linear input-varying systems, which generalises attention, convolutions, recurrence, and state-space layers as instances of a single mathematical family with different parameterisations. Gradient-free evolutionary operators including crossover, mutation, and tournament selection iterate over a population of these genomes, with each candidate trained briefly on a fixed corpus before being scored against the target metrics. The most promising architectures from each generation are mutated and recombined to produce the next.
The STAR framework matters because it allows Liquid AI to ship different architectures for different deployment targets without manually designing each one. The Hyena Edge model, demonstrated in April 2025, was a STAR-evolved architecture that replaced two-thirds of the Grouped-Query Attention layers in a baseline transformer with Hyena-Y gated convolutions and was benchmarked running on a Samsung Galaxy S24 Ultra smartphone.[22][23] The same approach scales to laptop and server form factors, and the architecture-search results that fed into LFM2 in mid-2025 were also produced by a STAR-style hardware-in-the-loop search.[19]
In April 2025, in the run-up to the International Conference on Learning Representations (ICLR), Liquid AI published Hyena Edge as a concrete proof point for STAR's edge claims.[22] Hyena Edge is a convolution-heavy hybrid model trained on 100 billion tokens and benchmarked against a parameter-matched Transformer++ baseline.[23] On the Samsung Galaxy S24 Ultra, Hyena Edge delivered prefill and decode latencies up to 30% faster than the transformer baseline at the same parameter count, with the speed advantage widening at longer sequence lengths, and it used less RAM during inference across every sequence length tested.[22][23] Liquid AI also reported competitive perplexity on Wikitext, Lambada, PiQA, HellaSwag, Winogrande, ARC-easy, and ARC-challenge.[23]
Hyena Edge mattered for two reasons. First, it was the first widely circulated demonstration of a STAR-evolved architecture beating a manually designed transformer on an actual consumer device rather than in a datacentre.[22] Second, it foreshadowed the LFM2 design philosophy that pairs short-range convolutional blocks with a smaller share of grouped-query attention, a recipe that became the backbone of the second-generation LFMs three months later.[5] The Hyena Edge weights were subsequently open-sourced as part of the company's pivot toward more transparent releases.
The second generation of Liquid Foundation Models, LFM2, launched on July 10, 2025 with a stronger emphasis on on-device inference.[5][18] LFM2 uses a hybrid block design with 16 layers per model: 10 double-gated short-range convolution blocks paired with 6 grouped-query attention blocks.[19] The company claims the new design delivers roughly 2x faster decode and prefill on CPU than Qwen3 of comparable size and a 3x improvement in training efficiency over the first LFM generation.[5] The LFM2 family was extended through late 2025 to include a 2.6B dense model and an 8.3B sparse mixture-of-experts variant, all with a 32k native context length.
| Model | Parameters | Type | Notes |
|---|---|---|---|
| LFM2-350M | 350 million | Dense | Smallest open-weight LFM2; phones and embedded |
| LFM2-700M | 700 million | Dense | Outperforms Gemma 3 1B IT on internal benchmarks |
| LFM2-1.2B | 1.2 billion | Dense | Competitive with Qwen3-1.7B at 47% fewer parameters[5] |
| LFM2-2.6B | 2.6 billion | Dense | Added in late 2025; targets laptops and high-end phones |
| LFM2-8B-A1B | 8.3 billion total, 1.5 billion active | Sparse Mixture-of-Experts | First MoE in the LFM2 family |
| LFM2-VL-450M / 1.6B / 3B | 0.45 / 1.6 / 3 billion | Vision-language | SigLIP2 NaFlex vision tower with LFM2 backbone[7] |
| LFM2-Audio | Not fully disclosed | Speech-language | Real-time speech-to-speech variant |
| LFM2-ColBERT | Not fully disclosed | Retrieval | Late-interaction encoder for multilingual search |
LFM2 weights are released under a Liquid Open License modelled on Apache 2.0 that permits free use for academic work and for companies with annual revenue under $10 million; larger commercial users need a separate licence.[18] The models are distributed through Hugging Face, the Liquid Playground, llama.cpp, and OpenRouter.[5] A late-2025 follow-up branded LFM2.5 carried the same hybrid architecture into a model family aimed at on-device agents.
The LFM2 technical report, released on arXiv in November 2025 as 2511.23404, is the most detailed disclosure the company has made.[19] It describes a hardware-in-the-loop architecture search that selects 10 double-gated short-range LIV convolution blocks and 6 grouped-query attention blocks per model, a 10 to 12 trillion token pretraining regime followed by a long-context mid-training phase, and a three-stage post-training recipe consisting of supervised fine-tuning, length-normalised preference optimisation, and model merging.[19] Knowledge distillation from a larger teacher model is applied through a tempered, decoupled top-K objective that the report argues avoids the support-mismatch problem associated with conventional logit distillation.[19]
The LFM2-VL multimodal models, released in August 2025 and extended in October 2025 with a 3B variant, pair a SigLIP2 NaFlex vision encoder with the LFM2 language backbone through a two-layer MLP projector with pixel unshuffle.[7] The vision-language stack was trained on roughly 100 billion multimodal tokens with a curriculum that started at 95% text and 5% image and ended at 30% text and 70% image.[7] Liquid AI claims up to 2x inference speedups over comparable open-weight vision-language models on GPU.[7] LFM2-Audio separates the audio input and output pathways to enable real-time speech-to-speech interaction; the report claims it is competitive with audio models roughly three times its size.[19] LFM2-ColBERT, a late-interaction retrieval encoder, is intended to act as the embedding side of a retrieval-augmented generation pipeline running on the same device as the LFM2 generator.
Liquid AI offers three deployment paths: a hosted API and chat playground at liquid.ai, on-premises inference packages for enterprise customers, and the Liquid Edge AI Platform (LEAP) for client devices.[25] LEAP launched in July 2025 alongside a developer app called Apollo[25] and was extended in August 2025 with native support for AMD Ryzen and Ryzen AI processors, including Zen 5 cores and integrated AMD Radeon graphics.[10] LEAP also runs on NVIDIA Jetson modules and on Apple Silicon, and Liquid AI has demonstrated LFM2 inference inside the browser via WebGPU.
LEAP is designed to be operating-system agnostic and model agnostic.[25] The initial release exposed an SDK for iOS and Android with bindings for Swift and Kotlin and an HTTP server mode for desktop platforms.[25] Liquid Apollo, originally an independent iOS app that the company acquired in early 2025 and rebuilt against the LEAP runtime, ships LFM2 inference inside a private chat interface that does not send conversation contents off the device.[25] The LEAP store within Apollo lets developers pull additional Liquid models or community fine-tunes without going through a system app store.
In August 2025, AMD and Liquid AI jointly announced LEAP support for AMD Ryzen and Ryzen AI processors running Zen 5 cores.[10] The announcement turned the strategic relationship from Series A into a concrete distribution play: any Windows or Linux laptop with a recent Ryzen part gains access to GPU-accelerated LFM2 inference through the LEAP runtime.[10] AMD and Liquid AI ran a joint developer hackathon, branded Hack The Edge, during the autumn of 2025 to seed an early developer community around the combination.
The company's commercial strategy leans heavily on the edge story. Where most foundation model providers compete on raw cloud capability, Liquid AI is positioning LFMs as the way to put a capable language or vision model on a phone, vehicle, household appliance, or industrial sensor without a network round trip. The April 2026 Mercedes-Benz announcement is the clearest example: Mercedes-Benz will run embedded LFMs on third- and fourth-generation MBUX systems in North America starting in the second half of 2026, handling speech, language understanding, and reasoning for the in-vehicle voice assistant without depending on the cloud.[11] The Mercedes deployment is built on the Mercedes-Benz Operating System (MB.OS) and complements rather than replaces the cloud LLMs that the manufacturer also uses.[11]
Liquid AI's partnership list weights heavily toward enterprise rather than consumer customers. The most consequential public partnerships as of early 2026 are:
| Partner | Announced | Nature |
|---|---|---|
| AMD | December 2024 | Series A lead investor; co-training on AMD Instinct GPUs; native LEAP support on Ryzen and Ryzen AI[6][10] |
| Capgemini | January 2024 | Strategic seed-round investor; go-to-market partner for enterprise Liquid Neural Network deployments[9] |
| Samsung Next | December 2023 | Seed-round investor; Hyena Edge demonstrated on Galaxy S24 Ultra in April 2025[21][22] |
| Shopify | November 2025 | Multi-year licensing deal; first Shopify production model powers sub-20ms search across merchant storefronts; HSTU recommender system co-developed with Liquid AI[8][20] |
| Mercedes-Benz | April 2026 | Multi-year partnership; embedded LFMs ship in third- and fourth-generation MBUX in North America from H2 2026[11] |
| Lambda Labs and Perplexity Labs | September 2024 | Hosted launch partners for the original LFM-1B/3B/40B release[15] |
The Shopify partnership, announced on November 13, 2025, is structurally interesting because it spans both equity and licensing.[8] Shopify participated in the December 2024 Series A and formalised a multi-year commercial agreement eleven months later, with Liquid AI building dedicated foundation models for product search, recommendations, classification, and agent workflows on Shopify's platform.[8] The first production deployment was a sub-20 millisecond text model serving search queries across merchant storefronts; the second was a generative recommender system based on the Hierarchical Sequential Transduction Unit (HSTU) architecture that the two companies say outperforms Shopify's previous recommendation stack on conversion rate in controlled tests.[8][20] Future workloads under evaluation include multimodal customer profiles and on-platform agents.
The Mercedes-Benz deal is the company's most prominent on-device customer reference outside the cloud and laptop tiers. Mercedes-Benz will embed LFMs in third- and fourth-generation MBUX, the infotainment platform built on MB.OS, starting with North American models in the second half of 2026.[11] The system handles speech recognition, natural language understanding, and reasoning entirely on the vehicle's compute, with cloud LLMs used in a complementary rather than primary role.[11] Independent press coverage emphasised the privacy and latency benefits of the on-vehicle approach, particularly for navigation, climate control, and entertainment queries that do not require external knowledge.
Reports of a Ricoh partnership appear in some secondary coverage but are not confirmed in primary Liquid AI announcements as of this writing.
Liquid AI is one of a small set of companies attempting to scale a non-transformer architecture to foundation model size. The field has settled into a few clusters: transformer-only labs (the largest commercial AI providers), pure state-space-model labs, and hybrid-architecture labs in which Liquid AI sits.
| Company | Founded | Architecture | Primary focus | Largest disclosed model | License |
|---|---|---|---|---|---|
| Liquid AI | 2023 | Hybrid Liquid + convolution + sparse attention | Edge and on-device foundation models | LFM-40B MoE / LFM2-8B-A1B | Liquid Open License (LFM2); proprietary (LFM1) |
| Cartesia | 2023 | Pure state-space (Mamba lineage) | Real-time speech and long-context | Sonic-2; Mamba-3 research | Apache 2.0 / proprietary |
| RWKV (open-source) | 2021 | Linear-attention RNN | Open foundation models | RWKV-7 | Apache 2.0 |
| Mistral AI | 2023 | Transformer + Mixture-of-Experts | Open and commercial transformer LLMs | Mistral Large 2 | Apache 2.0 / Mistral Research License |
| DeepSeek | 2023 | Transformer + MoE with multi-head latent attention | Open reasoning and coding LLMs | DeepSeek-V3 (671B) | DeepSeek License |
| Anthropic | 2021 | Transformer | Frontier safety-focused models | Claude family | Proprietary |
| Google DeepMind | 2014 / 2023 merger | Transformer | Multimodal frontier models | Gemini family | Proprietary |
Liquid AI's closest cousin in research terms is Cartesia, the state-space model startup founded by Albert Gu, Tri Dao, Karan Goel, Chris Re, and Brandon Yang. Both companies are commercializing continuous-time and selective recurrence as alternatives to attention. The architectural difference is that Cartesia's models are mostly pure state-space with selective scan operators, while Liquid AI's models are hybrid stacks that mix Liquid blocks with convolutions and a smaller share of attention. Both companies share a research lineage with Hyena and the broader linear-attention literature.
The broader competitive landscape includes purely open initiatives such as RWKV, which has been iterating publicly on linear-attention recurrent architectures since 2021, and conventional transformer labs that have begun experimenting with hybrid designs of their own. Within that landscape, Liquid AI's distinguishing assets are its specific Liquid block, the STAR search method that lets it tune the architecture per hardware target, and a commercial focus on edge deployment that very few other foundation model labs share.
A useful way to position Liquid AI against the dominant cloud labs is to look at where the cost is paid. Frontier transformer providers pay primarily in training compute and in datacentre inference economics, since their business is selling tokens generated on someone else's GPU fleet. Liquid AI also pays the training cost but trades datacentre inference economics for unit-shipment economics: every phone, laptop, vehicle, or appliance that runs an LFM is effectively prepaying for inference at manufacture time. That places the company in a different commercial weight class but creates different leverage, since a single design win at an automaker or device OEM can drive millions of recurring deployments per year.
The most consistent independent observation about LFMs is that they hit strong benchmark scores per parameter and per byte of memory. LFM2-1.2B beating Qwen3-1.7B on common benchmarks while running roughly twice as fast on CPU has been replicated by several third-party reviewers.[5][18] The company's edge story is also unusually credible because it is rooted in production partnerships with chip vendors and a major automaker rather than just demos.[10][11] The MIT pedigree of the founding team and the cleanly published academic foundation in the LTC and CfC papers give Liquid AI more architectural credibility than most non-transformer startups,[1][2] and the STAR architecture search lets the company target different hardware without throwing out the rest of the stack.[16]
A second strength is the maturing developer surface. LEAP, Apollo, and the Hugging Face releases under the Liquid Open License have brought the company much closer to the open ecosystems around Llama and Qwen than the closed initial LFM launch suggested.[25][18] The LFM2 technical report's level of detail, including pretraining token counts, post-training pipelines, and architecture-search procedures, is materially closer to the open-research standard set by DeepSeek and Mistral than to the disclosure level typical of frontier closed labs.[19]
Third, the LTC and CfC research lineage gives Liquid AI an interpretability angle that no transformer-only competitor can match.[2][17] While interpretability is not yet a primary commercial driver, the property is increasingly attractive in regulated industries such as automotive, healthcare, and finance, where being able to explain why a model made a given decision matters for certification and liability.
The original LFM-1B, LFM-3B, and LFM-40B weights were not open, and the company has provided fewer architectural details than is typical for a research-led foundation model lab. Independent benchmarks of long-context scaling and the more aggressive efficiency claims have been slower to appear than the launch-day numbers, and some reviewers found that LFM models trail the largest open-weight transformers on more demanding reasoning benchmarks even where they win on per-parameter efficiency. The multimodal stack, although growing fast, lags purpose-built vision-language models from larger labs, and the developer ecosystem around LFMs is much smaller than around Llama or Qwen.
The Liquid Open License's revenue cap also restricts the LFM2 weights' commercial reach in a way that pure Apache 2.0 releases avoid.[18] Companies with annual revenue above $10 million must negotiate a separate commercial agreement to deploy LFM2 in production, which has slowed adoption at firms that prefer to ship open-source software with predictable licensing.[18] By contrast, Mistral's and DeepSeek's permissive licensing has helped them seed enterprise adoption faster, even where the models themselves are arguably less efficient per parameter.
A more architectural concern is that the benchmark advantages claimed for LFMs are often measured against transformer baselines that are not themselves state of the art for the same parameter budget. Critics have pointed out that the LFM2 reports compare against open transformer baselines rather than against the most recent closed frontier models, and that the Hyena Edge headline numbers depend on a particular Transformer++ implementation that may not reflect the best edge-tuned transformers available.[23] These critiques do not undermine the central efficiency claim but do affect how the more aggressive marketing numbers should be read.
Finally, the scaling story for non-transformer architectures is still partially unproven. Liquid AI's largest publicly disclosed dense model is in the 2.6 billion parameter range, with the 8.3 billion parameter LFM2-8B-A1B being a sparse mixture-of-experts variant. Whether the hybrid Liquid plus convolution recipe continues to outperform pure attention at the hundreds-of-billions-of-parameter scale that frontier closed labs operate is still an open empirical question. The company's commercial bet does not require winning at that scale, but the long-run research bet partially does.
The rough timeline of Liquid AI's first three years runs as follows. March 2023, incorporation.[3] December 2023, $37.6 million seed round including Capgemini and a launch partnership at a roughly $303 million valuation.[21] September 30, 2024, public launch of LFM-1B, LFM-3B, and LFM-40B.[14][15] December 2024, $250 million Series A led by AMD Ventures at a roughly $2.3 billion valuation[6][12] and the unveiling of the STAR architecture search framework.[16] April 2025, demonstration of the Hyena Edge model on Samsung Galaxy S24 Ultra hardware.[22] July 2025, launch of LFM2 in 350M, 700M, and 1.2B sizes alongside the LEAP edge platform and Apollo developer app.[5][25] August 2025, LEAP support extended to AMD Ryzen and Ryzen AI laptops;[10] LFM2-VL multimodal models released.[7] October 2025, LFM2-VL-3B released.[7] November 2025, multi-year Shopify partnership announced, with the first Liquid-powered search model live on Shopify storefronts at sub-20 ms latency;[8] LFM2 Technical Report posted to arXiv.[19] Late 2025, LFM2.5 family released for on-device agents and the LFM2-2.6B dense model added to the family. April 2026, Mercedes-Benz announces a multi-year partnership to embed LFMs into MBUX in North America from the second half of 2026.[11]
Three threads run through that timeline. The first is a steady shift from closed-weight cloud models in 2024 toward openly licensed on-device models in 2025 and 2026, mediated by the Liquid Open License and the LEAP runtime. The second is the deepening relationship with AMD, which moved from financial backer at Series A to a co-engineering partner whose Ryzen and Instinct silicon is the reference target for the LEAP and cloud training stacks respectively.[6][10] The third is the company's growing portfolio of marquee customer references in regulated, latency-sensitive industries: commerce with Shopify, automotive with Mercedes-Benz, and enterprise IT with Capgemini.[8][11][9]
As of mid-2026 Liquid AI is not yet profitable and has not disclosed revenue, but the combination of two large customer deployments, a chip-vendor co-marketing relationship, and a maturing open-weight developer ecosystem positions the company as one of the most visible non-transformer foundation model labs operating at scale. Whether the bet on continuous-time Liquid units and architecture search holds up against the still-improving pure-transformer stack at the multi-trillion-parameter frontier remains the defining open question for the company's research roadmap.