# Liquid AI

> Source: https://aiwiki.ai/wiki/liquid_ai
> Updated: 2026-10-28
> Fact-checked: 2026-10-28
> Categories: AI Companies, AI Models, Model Architecture
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "Liquid AI." aiwiki.ai, 28 Oct 2026. https://aiwiki.ai/wiki/liquid_ai
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

**Liquid AI** is an [artificial intelligence](https://aiwiki.ai/wiki/artificial_intelligence) company founded in 2023 by researchers from MIT CSAIL and headquartered in Cambridge, Massachusetts.[3][63] The company develops Liquid Foundation Models (LFMs), a family of neural networks built on non-transformer and hybrid architectures that grew out of its founders' research on Liquid Time-Constant (LTC) networks.[4][1] LFMs are designed to need less memory than transformers of similar quality, particularly for long-context inference and on-device deployment.[15][35] In December 2024 it closed a $250 million Series A led by AMD, which Maginative reported valued it at $2.3 billion,[6][12][28] and since July 2025 it has released the LFM2 and LFM2.5 families of open-weight models built to run directly on phones, laptops, vehicles, and embedded systems,[5][43] followed on October 27, 2026 by the multimodal [LFM3](https://aiwiki.ai/wiki/lfm3) family.[78]

<figure class="w-56">
<img src="/images/site-2026-10-28-liquid_ai-logo-ef1a9df3a0f8-tile.png" alt="Liquid AI logo" width="261" height="108" />
<figcaption>Liquid AI logo, from the official website <a href="https://liquid.ai/">liquid.ai</a>. Trademark of its owner.</figcaption>
</figure>

The company was spun out of the [Massachusetts Institute of Technology](https://aiwiki.ai/wiki/mit) Computer Science and Artificial Intelligence Laboratory by Ramin Hasani, Mathias Lechner, Alexander Amini, and Daniela Rus.[3] Its public profile rose with the September 30, 2024 launch of the first three LFMs (LFM-1B, LFM-3B, and LFM-40B)[14][15] and the December 2024 Series A.[6][12] Since then it has announced a multi-year commerce partnership with [Shopify](https://aiwiki.ai/wiki/shopify) (November 2025),[8] an in-vehicle partnership with Mercedes-Benz (April 2026),[11] and an on-device context layer optimized for [Qualcomm](https://aiwiki.ai/wiki/qualcomm) Snapdragon processors (September 2026).[58] In October 2026 it released LFM2.5-24B-A2B, the largest and final model of the LFM2.5 family, and a new line of "decision models" called d1,[52][53] and on October 27 it announced LFM3, a third model generation built on what it calls a "tri-hybrid" architecture, with six open-weight models of about 2 billion parameters.[78] On stage at Qualcomm's Snapdragon Summit in September 2026, Hasani said Liquid AI's open-weight models were being downloaded about 1.4 million times a week on Hugging Face.[59]

Liquid AI's work on [foundation models](https://aiwiki.ai/wiki/foundation_models) combines biologically inspired [neural network](https://aiwiki.ai/wiki/neural_network) research with the search for transformer alternatives that use less memory on long inputs and the push to run capable models on end-user devices instead of in the cloud. The founders' early work drew on the nervous system of the nematode *Caenorhabditis elegans*, which has only 302 neurons,[17] and the company argues that compact, dynamically evolving computational units can deliver competitive quality with far smaller parameter counts and memory budgets, an advantage it says matters most on devices where latency, privacy, and memory constraints are tightest.[33][5]

## What is Liquid AI?

Liquid AI is a Cambridge, Massachusetts foundation model company that builds Liquid Foundation Models, a non-transformer class of [generative AI](https://aiwiki.ai/wiki/generative_ai) models, with a commercial focus on running language, vision, speech, and decision models on-device rather than in the cloud.[4][5] VentureBeat described the first LFMs as non-transformer models, and the company presents efficiency per parameter and per byte of memory as their main advantage.[15][5]

The company 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).[17] Hasani told TechCrunch that the idea was invented at TU Wien in Radu Grosu's lab, where he completed his PhD and Lechner his master's degree, and was then refined and scaled in Rus's lab at MIT CSAIL.[27] The most important precursor paper, "Liquid Time-constant Networks" by Hasani, Lechner, Amini, Rus, and Grosu, was posted to [arXiv](https://aiwiki.ai/wiki/arxiv) as 2006.04439 in June 2020 and presented at AAAI 2021.[1] A follow-up paper, "Closed-form continuous-time neural networks," appeared in *Nature Machine Intelligence* on November 15, 2022 and showed that the interaction between neurons and synapses in LTC networks could be closely approximated in closed form, removing the need for a numerical differential equation solver; the authors reported models between one and five orders of magnitude faster in training and inference than their differential-equation-based counterparts.[2]

The LTC architecture was partly inspired by *C. elegans*, a millimetre-long roundworm whose nervous system has been mapped in its entirety.[33][17] In a 2020 *Nature Machine Intelligence* paper, Lechner, Hasani, Amini, Thomas Henzinger, Rus, and Grosu reported that a controller with 19 control neurons, connecting 32 input features to its outputs through 253 synapses, learned to map high-dimensional camera inputs to steering commands for an autonomous vehicle, and that it was more generalizable, interpretable, and robust than orders-of-magnitude larger black-box models.[32] Science News reported that Rus learned of Hasani's graduate work on *C. elegans* at a time when she was training self-driving models with hundreds of thousands of artificial neurons and half a million parameters, and that she invited him and a colleague to move to MIT.[34]

## When was Liquid AI founded, and where is it based?

Liquid AI was founded in 2023 and emerged from stealth on December 6, 2023, announcing a two-stage seed round that TechCrunch put at $37.5 million and SiliconANGLE and the company put at $37.6 million.[27][21][9] TechCrunch reported that the tranche valued the company at $303 million post-money and named investors OSS Capital, PagsGroup, Automattic, Samsung Next, Bold Capital Partners, and ISAI Cap Venture, along with angel investors Tom Preston-Werner, Tobias Lütke, and Bob Young.[27] Capgemini participated through ISAI Cap Venture, its corporate venture fund managed by ISAI.[9] A Form D notice filed with the U.S. Securities and Exchange Commission on December 18, 2023 lists Liquid AI, Inc. as a Delaware corporation incorporated in 2023, reports $31.9 million of securities sold in an offering whose first sale was dated December 6, 2023, and names Hasani, Lechner, Rus, Stephen Pagliuca, and Joseph Jacks as directors.[26] At the time the company had a 12-person team and a presence in Boston and Palo Alto.[27]

Company press releases describe Liquid AI as headquartered in Cambridge, Massachusetts.[63][64] As of October 2026 its careers page listed roles in San Francisco, Boston, and Tokyo and said hiring was primarily in San Francisco and Boston, with a Tokyo presence.[69] AMD described the November 2025 Hack the Edge event as hosted at Liquid AI's office in San Francisco.[67]

## Who founded Liquid AI?

The four co-founders all worked in or with Daniela Rus's MIT CSAIL group; Hasani and Lechner came to MIT from TU Wien and ISTA.[3][17][27] TechCrunch listed Hasani as CEO, Lechner as CTO, and Amini as chief scientific officer when the company emerged from stealth in December 2023, and the company's About page lists the same roles as of October 2026.[27][3]

| Name | Role at Liquid AI | Background |
|---|---|---|
| Ramin Hasani | Co-founder and CEO | PhD completed with honors in May 2020 at TU Wien, in Radu Grosu's lab;[27][76] postdoctoral associate in the Distributed Robotics Lab at MIT CSAIL;[17][76] first author on the LTC and CfC papers;[1][2] TechCrunch reported he had been principal AI scientist at Vanguard[27] |
| Mathias Lechner | Co-founder and CTO | BSc (2016) and MSc (2017) in computer science, TU Wien; PhD (2022), IST Austria, supervised by Thomas Henzinger; research affiliate at MIT CSAIL[29] |
| Alexander Amini | Co-founder and Chief Scientific Officer | BSc (2017), MSc (2018), and PhD (2022) in computer science, MIT; creator and lead lecturer of MIT 6.S191 Introduction to [Deep Learning](https://aiwiki.ai/wiki/deep_learning); co-founder of Themis AI[30] |
| Daniela L. Rus | Co-founder and board director[26] | Director of MIT CSAIL since May 2012;[31] Andrew and Erna Viterbi Professor of Electrical Engineering and Computer Science;[17] PhD in computer science, Cornell University; MacArthur Fellow (2002)[31] |

Before Liquid AI, Hasani's research centered on small continuous-time networks for control and time-series problems, including the neural circuit policies work and the LTC and CfC papers.[32][1][2] Lechner and his ISTA colleague Djordje Zikelic received the institute's 2023 Outstanding Scientific Achievement Award for work on proving safety in stochastic machine learning systems.[29] Amini is also the lead lecturer of MIT 6.S191, whose lectures have drawn more than 10 million views according to his biography.[30]

## How much funding has Liquid AI raised?

Liquid AI has announced two financing rounds: a two-stage seed in 2023 and a $250 million Series A in December 2024.[27][6] AMD's involvement in the Series A doubled as a technical partnership: TechCrunch reported that, as part of the investment, Liquid AI would work with AMD to optimize its models for AMD's GPUs, CPUs, and AI accelerators.[12]

| Round | Date | Amount | Lead investor | Selected participants | Reported valuation |
|---|---|---|---|---|---|
| Seed (two stages) | December 2023 | $37.5 million (TechCrunch); $37.6 million (SiliconANGLE, company) | OSS Capital and PagsGroup (co-led) | Automattic, Samsung Next, Bold Capital Partners, ISAI Cap Venture (through which Capgemini invested), Tom Preston-Werner, Tobias Lütke, Bob Young | $303 million post-money for the tranche (TechCrunch)[27][21][9] |
| Series A | December 13, 2024 | $250 million | AMD | OSS Capital, Duke Capital Partners, PagsGroup, [Shopify](https://aiwiki.ai/wiki/shopify); Lechner's website also names Samsung and G42 | $2.3 billion (Maginative; also stated on Lechner's website); "over $2 billion" (TechCrunch, citing Bloomberg)[6][12][13][28][29][8] |

The seed cap table foreshadowed several later commercial relationships. Capgemini announced a collaboration with Liquid AI at CES in January 2024 to build AI solutions for enterprises in manufacturing, healthcare, and finance, saying it would play a primary role for Liquid AI in the market; the announcement said the relationship grew out of Capgemini's Strategic University Research Partnership Program with MIT and a two-year joint project that used liquid neural networks to train vehicle control systems.[9] Tobias Lütke, the Shopify co-founder, invested personally in the seed; Shopify itself participated in the Series A and signed a multi-year commercial agreement in November 2025.[27][8] Tom Preston-Werner, the [GitHub](https://aiwiki.ai/wiki/github) co-founder and an early investor, told WIRED in 2024 that "finding a new type of foundation model doesn't happen every day."[33] Bob Young is a co-founder of [Red Hat](https://aiwiki.ai/wiki/red_hat).[27]

The Series A was led by a chipmaker rather than a generalist venture firm.[12] In the blog post announcing the round, Liquid AI said it would "use our Series A funding to scale our compute infrastructure, expedite our product readiness on the edge and on-premise" and described LFMs as "our lightweight, general-purpose AI models that unlock private, efficient, and reliable enterprise-grade AI for all."[6] It said it would integrate its products into consumer electronics, telecommunications, financial services, e-commerce, and biotechnology, and scale LFMs across more model sizes and data modalities.[6] Mathew Hein, AMD's Senior Vice President and Chief Strategy Officer of Corporate Development, said in the same post that AMD was "thrilled to collaborate with Liquid AI to train and deploy their AI models on [AMD](https://aiwiki.ai/wiki/amd) Instinct GPUs."[6] Lechner's personal website describes the round as led by AMD, Shopify, Samsung, and G42 at a valuation of $2.3 billion.[29]

As of October 2026, Liquid AI's news pages listed no funding announcement after the Series A, and the company had not disclosed a new valuation.[71] The venture firm YXS Capital lists Liquid AI as a portfolio company on its website, without stating the amount, date, or terms of its investment.[70]

## What are Liquid Time-Constant networks?

Liquid Time-Constant (LTC) networks are continuous-time [recurrent neural networks](https://aiwiki.ai/wiki/recurrent_neural_network) 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 main 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 authors describe the formulation as loosely related to computational models of neural dynamics in small species: the potential of a non-spiking neuron can be written as a linear differential equation driven by its synaptic inputs and a leakage conductance, and approximating the synaptic currents with a sigmoidal nonlinearity yields an equation similar to the LTC form.[1]

The LTC paper shows that LTCs are universal approximators and that they exhibit stable and bounded behavior.[1] The original implementation computed the network's state with a fused numerical ODE solver and trained it with [backpropagation](https://aiwiki.ai/wiki/backpropagation) through time, which was expensive at training time; across eleven time-series prediction experiments, the paper reported improved performance over modern recurrent networks in a majority of cases.[1] The 2022 Closed-form Continuous-time (CfC) paper showed that the LTC dynamics can be approximated in closed form, so the same kind of model can be evaluated with a normal forward pass and no ODE solver.[2]

The CfC paper reported that closed-form models performed well in [time series](https://aiwiki.ai/wiki/time_series) modelling compared with advanced recurrent neural network models while running one to five orders of magnitude faster than ODE-based counterparts.[2] Compactness of this kind, demonstrated on control and time-series tasks, is the property Liquid AI set out to scale up to natural language, vision, and speech.[33]

Another property that motivates the company's edge focus is interpretability. MIT News reported in 2021 that because Hasani's network used a small number of highly expressive neurons, it was easier to peer into the "black box" of its decision making,[17] and WIRED reported that the behavior of liquid networks "can essentially be rewound to see how it produced an output."[33] Interpretability was a selling point of the original liquid networks, although the company's LFM2-, LFM2.5-, and LFM3-family [language model](https://aiwiki.ai/wiki/language_model)s are hybrids of gated short convolutions and attention rather than LTC networks.[19][78] The architecture search behind LFM2 included the founders' CfC networks among its candidate blocks, alongside state-space and linear-attention variants, but repeatedly selected the simpler convolution-and-attention hybrid.[19]

## What are Liquid Foundation Models (LFMs)?

The first generation of Liquid Foundation Models was released as a preview on September 30, 2024.[35][14] The launch covered three models intended to span on-device, edge, and cloud workloads.

| Model | Parameters | Architecture | Position |
|---|---|---|---|
| LFM-1B | 1.3 billion | Dense | Highly resource-constrained environments[35] |
| LFM-3B | 3.1 billion | Dense | Edge deployment, such as mobile applications, robots, and drones[35][14] |
| LFM-40B | 40.3 billion total, 12 billion active | Mixture-of-Experts | More complex tasks, deployed on cloud servers[35][14] |

Liquid AI described the models as built from "computational units deeply rooted in the theory of dynamical systems, signal processing, and numerical linear algebra" and did not publish their detailed architecture at launch.[35][15] It said the models offered a best-in-class effective [context window](https://aiwiki.ai/wiki/context_window) of 32,000 tokens, as measured on the RULER benchmark, and that their inference memory grows much more slowly with input length than that of transformer models, whose KV cache grows linearly with sequence length.[35] VentureBeat and SiliconANGLE reported the company's claim that the models could handle up to 1 million tokens efficiently with minimal impact on memory.[15][14]

On benchmarks reported at launch, Liquid AI said LFM-1B achieved the highest scores across various benchmarks in the 1B category, using EleutherAI's lm-evaluation-harness.[35] It said LFM-3B outperformed the previous generation of 7B and 13B models and was on par with [Microsoft](https://aiwiki.ai/wiki/microsoft)'s [Phi](https://aiwiki.ai/wiki/phi)-3.5-mini on multiple benchmarks while being 18.4% smaller.[35] LFM-40B, which activates 12 billion parameters per token, scored 78.76 on five-shot [MMLU](https://aiwiki.ai/wiki/mmlu) in a comparison table that also included larger models such as Meta's [Llama](https://aiwiki.ai/wiki/llama) 3.1 70B.[35] WIRED reported the company's claim that LFM-40B outperformed the 70-billion-parameter Llama 3.1 on MMLU-Pro.[33] [Sébastien Bubeck](https://aiwiki.ai/wiki/sebastien_bubeck), then a researcher at OpenAI, told WIRED that "the benchmark results for their SLMs look very promising."[33]

The models were offered in early access through the Liquid Playground, [Lambda Labs](https://aiwiki.ai/wiki/lambda_labs) (through Lambda Chat and its API), and [Perplexity](https://aiwiki.ai/wiki/perplexity) Labs, and Liquid AI said it was optimizing them for NVIDIA, AMD, Apple, Qualcomm, and Cerebras hardware.[14] Weights for the first generation of LFMs were not released openly.[15] At an event at MIT on October 23, 2024, the company also showed models for detecting fraud in financial transactions, controlling self-driving cars, and analyzing genetic data.[33] The later LFM2 line was released with open weights and accompanied by a technical report.[5][19]

In January 2025 the company added LFM-7B, a 7-billion-parameter model that it described as best-in-class for chat in English, Arabic, and Japanese and offered through the Liquid Playground and AWS Marketplace.[36]

## How does STAR architecture search work?

In late November 2024, Liquid AI researchers Armin W. Thomas, Rom Parnichkun, Alexander Amini, Stefano Massaroli, and Michael Poli published STAR, short for Synthesis of Tailored Architectures, a framework for automatically designing model architectures.[37][16][24] STAR uses an evolutionary search over a hierarchical numerical encoding the company calls "STAR genomes." Its search space is based on the theory of linear input-varying (LIV) systems and covers attention, recurrences, and convolutions.[37][16][24] The search optimizes several quality and efficiency metrics at once, such as model quality, parameter count, and inference cache size.[37] According to VentureBeat, STAR-evolved architectures achieved cache size reductions of up to 37% compared with hybrid models and 90% compared with transformers while maintaining or exceeding their predictive performance, and reduced parameter counts by up to 13% while improving results on standard benchmarks.[16]

Under the hood, STAR represents a candidate model as a numeric sequence whose entries encode both the computational units used and the connections between them.[37][24] An LIV operator is a linear operator whose weights are generated on the fly from its input, which lets convolutions, recurrences, attention, and other structured layers be described in one framework.[5][37] Gradient-free evolutionary operators, namely tournament selection, recombination, and mutation, iterate over a population of these genomes; in the paper's experiments, candidate models were trained from scratch on 1.3 billion tokens before being scored.[37]

STAR lets Liquid AI search for different architectures for different deployment targets without designing each one by hand. The Hyena Edge model, demonstrated in April 2025, was a STAR-evolved architecture,[22][23] and LFM2 was produced by a modified STAR search that used more than 50 internal evaluations and measured peak memory and prefill and decode speed directly on Qualcomm Snapdragon embedded CPUs.[5] For LFM3 the same kind of hardware-in-the-loop search was widened to 14 device and backend combinations.[78]

Three of the STAR co-authors later left to co-found [Radical Numerics](https://aiwiki.ai/wiki/radical_numerics), a biology-focused AI lab that Fortune reported raised a $50 million seed round in June 2026: Poli became its chief AI scientist, Massaroli its president, and Thomas its chief technology officer.[83] The Radical Numerics website describes Poli and Massaroli as founding scientists at Liquid AI and Thomas as a former senior researcher on automated architecture design there.[82]

## What is Hyena Edge, and how did the smartphone demonstration go?

In April 2025, in the run-up to the [International Conference on Learning Representations](https://aiwiki.ai/wiki/iclr) (ICLR) in Singapore, Liquid AI published Hyena Edge, an architecture designed with STAR for edge devices.[22][23] The search started from a population of 16 candidate architectures evolved over 24 generations, drawing on 18 convolution operators from the [Hyena](https://aiwiki.ai/wiki/hyena) family (Hyena, Hyena-X, and Hyena-Y with varying filter lengths) as well as attention and SwiGLU variants.[23] The final design replaces two-thirds of the [Grouped-Query Attention](https://aiwiki.ai/wiki/grouped_query_attention) operators in a GQA-Transformer++ baseline with gated convolutions from the Hyena-Y family.[23] Hyena Edge and the parameter-matched baseline were both trained on the same 100 billion tokens.[23] On a Samsung Galaxy S24 Ultra, Hyena Edge delivered up to 30% faster prefill and decode latencies at longer sequences, faster prefill even at the shortest sequence lengths, and lower memory use than the baseline at all sequence lengths.[22][23] Liquid AI also reported that it outperformed the baseline on Wikitext, Lambada, PiQA, [HellaSwag](https://aiwiki.ai/wiki/hellaswag), Winogrande, ARC-easy, and ARC-challenge.[22][23]

Hyena Edge was a STAR-evolved architecture measured against a transformer baseline on a consumer phone rather than in a data center.[22][23] Its mix of gated convolutions and a minority of grouped-query attention resembles the later LFM2 design, which pairs short-range gated convolutions with a small share of grouped-query attention.[5] When it announced Hyena Edge, Liquid AI said it planned to open-source a series of Liquid foundation models in the following months, including Hyena Edge;[23] the open-weight LFM2 models followed in July 2025.[5]

## What is LFM2, and why is it built for on-device use?

The second generation of Liquid Foundation Models, LFM2, launched on July 10, 2025 with open weights for three dense checkpoints of 350 million, 700 million, and 1.2 billion parameters.[5][18] Liquid AI describes LFM2 as a hybrid Liquid model with multiplicative gates and short convolutions; each of the three launch models has 16 blocks, of which 10 are double-gated short-range convolution blocks and 6 are grouped-query attention blocks.[5][19] The company claims the design delivers 2x faster decode and prefill on CPU than [Qwen](https://aiwiki.ai/wiki/qwen)3 and a 3x improvement in training efficiency over the first LFM generation.[5] The launch models were trained on 10 trillion tokens of roughly 75% English, 20% multilingual, and 5% code data, with the company's first-generation LFM1-7B model used as a teacher for knowledge distillation throughout pretraining, and with the context length extended to 32,000 tokens.[5] The family was extended in September and October 2025 with a 2.6B dense model and an 8.3B-parameter sparse [mixture of experts](https://aiwiki.ai/wiki/mixture_of_experts) model that activates 1.5B parameters per token.[38][39]

| Model | Parameters | Type | Notes |
|---|---|---|---|
| LFM2-350M | 350 million | Dense (16 blocks, 6 attention) | Competitive with Qwen3-0.6B and Llama 3.2 1B Instruct, per the company[5][19] |
| LFM2-700M | 700 million | Dense (16 blocks, 6 attention) | Outperforms Gemma 3 1B IT on the company's benchmarks[5] |
| LFM2-1.2B | 1.2 billion | Dense (16 blocks, 6 attention) | Competitive with Qwen3-1.7B, a model with a 47% bigger parameter count[5] |
| LFM2-2.6B | 2.6 billion | Dense (30 layers, 8 attention) | Released September 23, 2025; 82.41% on GSM8K and 79.56% on IFEval, per the company[38][19] |
| LFM2-8B-A1B | 8.3 billion total, 1.5 billion active | Sparse Mixture-of-Experts (32 experts, top-4) | Released October 7, 2025; first MoE in the LFM2 family[39][19] |
| LFM2-VL-450M / 1.6B / 3B | 0.45 / 1.6 / 3 billion | Vision-language | SigLIP2 NaFlex vision encoder with LFM2 backbone; 3B added October 22, 2025[7][74] |
| LFM2-Audio-1.5B | 1.5 billion | Audio-text | End-to-end speech and text model, released October 1, 2025[40] |
| LFM2-ColBERT-350M | 350 million | Retrieval | Multilingual late-interaction retriever, released October 28, 2025[41] |
| Liquid Nanos | 350 million to 2.6 billion | Task-specific | Extraction, English-Japanese translation, RAG, tool use, and math variants, first released September 25, 2025[42] |

LFM2 weights are released under the LFM Open License, which is based on Apache 2.0 and permits free academic and research use and commercial use by companies with less than $10 million in annual revenue; larger companies need a separate commercial licence.[7][38] As of October 2026, Liquid AI's models page said every LFM is free to download, run, and fine-tune, including commercially, until a company passes $10 million in annual revenue.[4] The models are distributed through [Hugging Face](https://aiwiki.ai/wiki/hugging_face), the Liquid Playground, [llama.cpp](https://aiwiki.ai/wiki/llama_cpp), and [OpenRouter](https://aiwiki.ai/wiki/openrouter).[5] Liquid AI said in February 2026 that the LFM2 family had passed 10 million downloads on Hugging Face.[45]

The LFM2 technical report was posted to arXiv on November 28, 2025 as 2511.23404; its 33 authors include all four co-founders.[19] It describes a hardware-in-the-loop architecture search under edge latency and memory constraints that repeatedly selected a minimal hybrid of mostly gated short convolution blocks with a small minority of grouped-query attention blocks; under the same on-device budgets, adding linear-attention, state-space, or extra convolution operators did not improve quality.[19] The report covers a 10 to 12 trillion token pretraining phase followed by a long-context mid-training phase, and a three-stage post-training recipe consisting of [supervised fine-tuning](https://aiwiki.ai/wiki/supervised_fine-tuning), length-normalised preference optimisation, and [model merging](https://aiwiki.ai/wiki/model_merging).[19] [Knowledge distillation](https://aiwiki.ai/wiki/knowledge_distillation) from a larger teacher model is applied through a tempered, decoupled top-K objective that the report says avoids support mismatch.[19]

The LFM2-VL [multimodal model](https://aiwiki.ai/wiki/multimodal_model)s, released on August 12, 2025 and extended on October 22, 2025 with a 3B variant built on LFM2-2.6B, pair a SigLIP2 NaFlex [vision encoder](https://aiwiki.ai/wiki/vision_encoder) (an 86M-parameter base encoder for the 450M model and a 400M shape-optimized encoder for the larger models) with the LFM2 language backbone through a two-layer MLP projector with pixel unshuffle.[7][74] The vision-language stack was trained on the order of 100 billion multimodal tokens, with the share of text data gradually reduced from 95% to 30% during joint mid-training.[7] Liquid AI claims up to 2x inference speedups over comparable open-weight vision-language models on GPU.[7] LFM2-Audio extends LFM2-1.2B to take continuous audio features as input and generate discrete audio tokens as output; the technical report says this separation enables real-time speech-to-speech interaction competitive with [audio models](https://aiwiki.ai/wiki/audio_models) three times larger.[40][19] LFM2-ColBERT-350M can store documents in one language and retrieve them with queries in other languages, and Liquid AI pitches it as a drop-in component for [retrieval-augmented generation](https://aiwiki.ai/wiki/retrieval_augmented_generation) pipelines.[41]

## What is LFM2.5?

LFM2.5 is the next generation of Liquid's models, built on the LFM2 architecture with more pretraining data and large-scale reinforcement learning. Liquid AI introduced it on January 5, 2026 with the LFM2.5-1.2B family (Base, Instruct, a Japanese chat model, the LFM2.5-VL-1.6B vision-language model, and the LFM2.5-Audio-1.5B audio-language model), extending pretraining from 10 trillion to 28 trillion tokens.[43] AMD and Nexa AI were launch partners for versions optimized for neural processing units.[43] The family grew through 2026 and was completed in October with LFM2.5-24B-A2B, which the company called "the largest and final model in the LFM2.5 family."[52]

| Date | Release | Notes |
|---|---|---|
| January 5, 2026 | LFM2.5-1.2B (Base, Instruct, JP), LFM2.5-VL-1.6B, LFM2.5-Audio-1.5B | Pretraining extended from 10T to 28T tokens; new audio detokenizer 8x faster than LFM2's Mimi detokenizer at the same precision on a mobile CPU[43] |
| January 20, 2026 | LFM2.5-1.2B-Thinking | Reasoning model that fits within 900 MB of memory on a phone[44] |
| February 24, 2026 | LFM2-24B-A2B (early checkpoint) | 24B total and about 2.3B active parameters; 40 layers, 64 experts per MoE block, top-4 routing; designed to fit in 32 GB of RAM; 17T tokens of pretraining at release[45] |
| March 31, 2026 | LFM2.5-350M | Pretraining extended from 10T to 28T tokens[46] |
| April 8, 2026 | LFM2.5-VL-450M | Updated small vision-language model[75] |
| May 28, 2026 | LFM2.5-8B-A1B | Context extended from 32,768 to 128,000 tokens; pretraining scaled from 12T to 38T tokens; vocabulary expanded from 65,536 to 128,000 tokens[47] |
| June 25, 2026 | LFM2.5-230M | Smallest model to date; 19T tokens of pretraining; 213 tokens per second decode on a Galaxy S25 Ultra and 42 on a Raspberry Pi 5, per the company[48] |
| August 4, 2026 | LFM2.5-2.6B | Agentic model; about 34T tokens of pretraining; vocabulary doubled to 128K; dedicated 128K context-extension phase[49] |
| August 12, 2026 | LFM2.5-VL-3B | Improved screen understanding, grounding, function calling, and multi-image input over LFM2-VL-3B[50] |
| August 20, 2026 | LFM2.5 DSpark drafts | Draft models for [speculative decoding](https://aiwiki.ai/wiki/speculative_decoding) of the 1.2B, 2.6B, and 8B-A1B models; up to 3.18x throughput on GPU and 2.87x on-device[51] |
| October 13, 2026 | LFM2.5-24B-A2B | Reasoning-only model; pretraining scaled from 17T to 33T tokens; context window extended to 131,072 tokens; vocabulary doubled from 65K to 128K[52] |

Other 2026 releases included LFM2.5 retrieval and encoder models in June and July.[75] As of October 2026, Liquid AI's models page listed the original LFM2 text checkpoints as deprecated.[4] Liquid AI said LFM2.5-24B-A2B was competitive with Qwen3.6-35B-A3B, a model with about 50% more total and active parameters, on five of eight agentic and tool-use benchmarks, while larger models kept an edge on knowledge-heavy benchmarks such as GPQA Diamond and on coding benchmarks such as SWE-Bench Verified.[52] It reported decode speeds of 184 tokens per second on an Apple M5 Max and 114 on an AMD Ryzen AI Max+ 395, and more than 21,000 output tokens per second on a single NVIDIA B300 GPU.[52] The model's post-training ended with agentic reinforcement learning inside agent harnesses such as [OpenClaw](https://aiwiki.ai/wiki/openclaw) and [Hermes Agent](https://aiwiki.ai/wiki/hermes_agent).[52] During an earlier teacher-specialization stage the company observed reward hacking: the model searched a repository's git history and sites such as GitHub, PyPI, and Hugging Face for answers and even recognized that it was running in a sandbox, after which Liquid AI tightened its sandboxes and reported results from sealed evaluation environments.[52] As a demonstration of the smallest model, Liquid AI ran LFM2.5-230M on a [Unitree](https://aiwiki.ai/wiki/unitree) G1 humanoid robot's onboard [NVIDIA Jetson](https://aiwiki.ai/wiki/nvidia_jetson) Orin, where it turned natural-language instructions into sequences of tool calls to pre-trained motion skills.[48]

## What are Liquid AI's decision models?

On October 5, 2026 Liquid AI introduced d1, which it calls a decision model: a model that reads unstructured text and images plus one or more questions in a single forward pass and returns a probability for each possible answer, without generating tokens.[53] It handles three question types (a yes/no question, a choice of one label from several, and a position on a scale), and Liquid AI says a text decision takes 200 to 300 milliseconds.[53][55] The company benchmarked d1 against [Jev](https://aiwiki.ai/wiki/jev), another model that answers decision questions, and said an experimental release of d1 the previous week had become the first model to rival Jev on text decisions.[53] On October 7 it released two open-weight versions, d1-3B, trained from LFM2.5-VL-3B, and d1-omni-600M, trained from the bidirectional LFM2.5-Encoder-350M with added audio and vision encoders.[54] On October 15 it launched d1-ultra on its API at $0.042 per million input tokens; unlike the earlier d1 models, d1-ultra is based on "a larger open-weight model with a series of Liquid-specific heavy customizations," which the company did not name. In a comparison across 112 evaluations, Liquid AI said d1-ultra outperformed Jev 1.14 on all tasks and gave better-calibrated probabilities.[55] On October 27 the company released two decision models built on LFM3, d1-2B and d1-Omni-2B, and said d1-2B replaces d1-3B with a third fewer parameters.[78]

## What is LFM3?

On October 27, 2026 Liquid AI announced [LFM3](https://aiwiki.ai/wiki/lfm3), the third generation of Liquid Foundation Models.[78] The first release consists of six open-weight models in the 2-billion-parameter class, all branched from a single LFM3-1.7B-Base backbone and covering text, vision, audio, combined vision-and-audio ("omni") input, and decision tasks.[78][80] Liquid AI says every model in the family can be fine-tuned on a single GPU, and it pitches the family as a starting point for "specialized intelligence": small models that customers adapt to one narrow task and run on their own devices.[78] The weights are released under the LFM Open License v1.0, the license with a $10 million annual revenue threshold used for earlier LFMs.[79][80]

The architecture keeps LFM2's gated short convolutions but changes the layer mix, which Liquid AI again chose with its hardware-in-the-loop architecture search, this time measured on 14 device and backend combinations: Samsung Galaxy S26 Ultra and OnePlus 15 phones, a Raspberry Pi 5, Apple M4 Max and AMD Ryzen AI laptops, Arm and x86 cloud CPUs, GPUs, and the Qualcomm Hexagon NPU.[78] The company calls the result a "tri-hybrid": of LFM3-1.7B's 24 layers, 13 are gated short convolutions (with a kernel size of 7, up from 3 in LFM2), 5 are [sliding-window attention](https://aiwiki.ai/wiki/sliding_window_attention) layers with a 256-token window, and 6 are global attention layers.[78][79] Other changes include a sigmoid gate that scales each attention head's output, partial [rotary position embeddings](https://aiwiki.ai/wiki/rotary_position_embedding) in the global attention layers, and normalization of both the input and the output of each block.[78] LFM3-1.7B-Base was pre-trained on 22 trillion tokens; the last 3.5 trillion were a mid-training stage with a larger share of agentic traces and tool calls, which also extended the context window to 256K tokens.[78]

| Model | Parameters | Modality | Notes |
|---|---|---|---|
| LFM3-1.7B | 1.73 billion | Text | 256K-token context; post-trained for agent harnesses such as [Hermes Agent](https://aiwiki.ai/wiki/hermes_agent), Pi, and OpenClaw; about 1 GB of memory at 4-bit precision[78][79] |
| LFM3-VL-2B | 1.92 billion | Vision-language | First LFM vision model to take video input; new in-house 164M-parameter vision encoder that Liquid AI says uses 3x less compute than the SigLIP2 encoder of earlier LFM-VL models; 128K context[78][80] |
| LFM3-Audio-2B | About 2 billion | Audio and text | Text or audio in, text and/or speech out; speech recognition, synthesis, and translation in 18 languages; tool calling; in-house causal audio encoder and LiquidCodec audio tokenizer[78] |
| LFM3-Omni-2B | 2.2 billion | Vision, audio, and text | Liquid AI's first omni model; spoken visual question answering and audio-visual video understanding[78][80] |
| d1-2B | 1.92 billion[84] | Decision (text, images, and video)[84] | Built on LFM3-VL-2B; replaces d1-3B; adds typed answers such as bounding boxes, extracted fields, and numeric estimates[78] |
| d1-Omni-2B | 2B class | Decision (adds audio input) | Built on LFM3-Omni-2B[78] |

The benchmark results published with the launch are Liquid AI's own evaluations. It said LFM3-1.7B outperformed LFM2.5-2.6B while using 35% fewer parameters, and that across a table of 29 agent, tool-use, coding, instruction-following, knowledge, and long-context benchmarks it averaged 49.37, against 47.24 for [Qwen3.5](https://aiwiki.ai/wiki/qwen3_5)-9B and 39.27 for Qwen3.5-4B.[78][80] The company acknowledged that larger models still won on agentic coding and knowledge tasks, and the model card says LFM3-1.7B is not recommended for agentic coding or complex multi-step reasoning because of its small size.[78][79] For LFM3-VL-2B, Liquid AI reported an average of 71.4 across 62 vision benchmarks, against 71.7 for [Gemma 4](https://aiwiki.ai/wiki/gemma_4) 12B, 70.0 for Gemini 3.5 Flash Lite, and 73.4 for GPT-5.6 Luna.[78] The announcement also described research fine-tunes that turned LFM3-VL-2B into a robot tool orchestrator, a vision-language navigation model, and a [vision-language-action model](https://aiwiki.ai/wiki/vision_language_action_model) for manipulation.[78]

LFM3 launched on Hugging Face, the Liquid Playground, and the Liquid AI API, with native support in llama.cpp, MLX, vLLM, and SGLang.[78] Liquid AI listed Ambarella, AMD, Arm, Intel (via OpenVINO), NVIDIA, OpenRouter, Qualcomm (via Qualcomm AI Hub and GenieX), and Vercel as day-zero launch partners, and said Mercedes-Benz and Shopify had partnered on pre-release testing and evaluation.[78]

## How does Liquid AI deploy its models, and what is the LEAP platform?

As of October 2026 Liquid AI offered several deployment paths: a hosted API through Liquid Console with an OpenAI-compatible endpoint; local builds in GGUF, MLX, and ONNX formats for tools such as [Ollama](https://aiwiki.ai/wiki/ollama) and [LM Studio](https://aiwiki.ai/wiki/lmstudio); GPU serving through [vLLM](https://aiwiki.ai/wiki/vllm), [SGLang](https://aiwiki.ai/wiki/sglang), and hosting platforms such as Baseten and Modal; and the Liquid Edge AI Platform (LEAP) for mobile apps.[4]

LEAP launched on July 15, 2025 alongside Liquid Apollo, a fully private AI chat app that Liquid AI said it had acquired earlier that year and that now had native support for all LEAP models.[25] LEAP is designed to be operating-system agnostic and model agnostic.[25] At launch it offered SDKs for Android and iOS, a library of small models including LFM2, model sizes as small as 300MB, and support for devices with 4GB or more of RAM.[25] Apollo is available on iOS and Android.[25]

On August 18, 2025, Liquid AI announced native LEAP support for AMD Ryzen and Ryzen AI processors with "Zen 5" cores and integrated AMD Radeon graphics, using llama.cpp as the inference engine; after benchmarking consumer hardware, the company said AMD Ryzen AI processors delivered the best performance for LEAP deployments.[10] AMD CTO Mark Papermaster said the pairing gave developers "sub-100-millisecond responsiveness on-device while keeping data on the user's PC."[10] In November 2025 AMD and Liquid AI co-hosted Hack the Edge, a 48-hour hackathon at Liquid AI's San Francisco office where more than 70 developers built projects on LFMs running on ROCm-accelerated AMD hardware.[67]

At Qualcomm's Snapdragon Summit 2026 in Maui, Liquid AI announced on September 23, 2026 that Liquid Context, an on-device context layer that builds a persistent understanding of a user from device signals (with the user's permission) and makes it available to AI agents, had been optimized for the Hexagon NPU in Snapdragon processors.[58][59] It also offered device makers Liquid Agent, an embedded agent powered by LFM2.5-2.6B.[58] Qualcomm CEO Cristiano Amon invited Hasani on stage during his day-two presentation, where Hasani said Liquid's open-weight models ranged from 200 million to 24 billion parameters.[59]

In August 2026 Liquid AI released Pipette, an open-source platform for benchmarking models on edge devices built in partnership with [Artificial Analysis](https://aiwiki.ai/wiki/artificial_analysis), with a public dataset of more than 1,000 model, quantization, runtime, device, and context configurations.[57] On October 22, 2026 it announced RIPL (Real-time Inference Pipeline), an inference engine for enterprise LFM deployments on NVIDIA GPUs that it says returns complete generated responses end-to-end in under 50 milliseconds while traffic arrives continuously at peak load.[56]

The company's commercial strategy centers on the edge. Liquid AI argues that moving generative models from the cloud to devices such as phones, laptops, cars, robots, and wearables unlocks millisecond latency, on-device resilience, and data-sovereign privacy.[5] The April 2026 Mercedes-Benz announcement applies this to cars: built on the Mercedes-Benz Operating System (MB.OS), the collaboration creates a path to a first production deployment of advanced speech technology for Mercedes-Benz models in North America, targeting the second half of 2026, and complements the cloud-based LLMs the carmaker also uses.[11]

## Who are Liquid AI's customers and partners?

Liquid AI's partners are mostly enterprises and device makers. Publicly announced partnerships and investor relationships as of October 2026 include:

| Partner | Announced | Nature |
|---|---|---|
| Samsung Next | December 2023 | Seed-round investor;[27] WIRED reported in October 2024 that Samsung was testing Liquid AI's technology[33] |
| Capgemini | January 2024 | Seed investor through ISAI Cap Venture; go-to-market collaboration for enterprise AI solutions[9] |
| [Lambda Labs](https://aiwiki.ai/wiki/lambda_labs) and Perplexity Labs | September 2024 | Early-access hosts for the original LFM-1B, LFM-3B, and LFM-40B release[14] |
| AMD | December 2024 | Series A lead; models trained and deployed on AMD Instinct GPUs; native LEAP support on Ryzen and Ryzen AI[6][10] |
| Sharp | June 2025 | Selected a 1B+1B vision-language model for its AQUOS wish4 smartphone[62] |
| [G42](https://aiwiki.ai/wiki/g42) | June 2025 | Commercial partnership, involving Core42 and Inception, to develop and deploy private generative AI solutions internationally[63] |
| Alef Education | August 2025 | Deployment of private, on-device generative models across Alef's education platform[64] |
| [Brilliant Labs](https://aiwiki.ai/wiki/brilliant_labs) | September 2025 | License for current and upcoming multimodal LFMs; Halo AI glasses use LFM2-VL-450M[65] |
| [Shopify](https://aiwiki.ai/wiki/shopify) | November 2025 | Multi-year licensing deal; sub-20ms search model in production; co-developed HSTU [recommender system](https://aiwiki.ai/wiki/recommender_system)[8][20] |
| [Insilico Medicine](https://aiwiki.ai/wiki/insilico_medicine) | March 2026 | Co-developed LFM2-2.6B-MMAI, a drug-discovery model[61] |
| Mercedes-Benz | April 2026 | Multi-year partnership; embedded LFMs for MBUX in North America, with a first production deployment targeted for the second half of 2026[11][66] |
| MacPaw | August 2026 | Long-term partnership on local AI for the Mac, starting with the Eney assistant[60] |
| Qualcomm | September 2026 | Liquid Context and Liquid Agent optimized for the Snapdragon Hexagon NPU[58][59] |
| Ambarella | October 2026 | LFM3 launch partner; Ambarella said developers could already run LFM2.5-VL models on its chips through its Cooper Model Garden and that it had begun enabling LFM3[78] |

The Shopify partnership, announced on November 13, 2025, spans both equity and licensing.[8] Shopify participated in the December 2024 Series A, and the multi-year, multi-million-dollar agreement formalized co-development already under way; financial terms were not disclosed.[8] The first production deployment was a sub-20 millisecond text model that enhances search; the companies also co-developed a generative recommender system with an HSTU architecture that they said outperformed Shopify's prior stack in controlled testing, producing higher conversion rates from recommendations.[8][20] The companies said they were evaluating multimodal models for customer profiles, agents, and product classification.[8] Shopify CTO Mikhail Parakhin said in the announcement: "I've seen a lot of models. No one else is delivering sub-20ms inference on real workloads like this."[8] In October 2026 Liquid AI cited Shopify's production use as an example for its RIPL inference engine.[56]

The Mercedes-Benz deal was announced on April 23, 2026. It covers Mercedes-Benz models with third- and fourth-generation MBUX in North America and evolves the MBUX Virtual Assistant by integrating voice control, vehicle functions, and contextual understanding.[11] Built on MB.OS, the collaboration "creates a path to a first production deployment for advanced speech technology, targeting the second half of 2026," and Mercedes-Benz CTO Jörg Burzer described "a clear path toward initial production deployment as early as the second half of 2026."[11][66] Burzer described the work as advancing on-device speech, [natural language understanding](https://aiwiki.ai/wiki/natural_language_understanding), and reasoning; the announcement said Liquid AI's models support conversational interaction without relying on continuous data exchange with the cloud, and that the partnership complements the cloud-based large language models the carmaker already uses.[11]

Other agreements cover devices and science. Liquid AI said Sharp chose its 1B+1B vision-language model after other open-source models such as Llama failed to meet Sharp's latency and memory needs for mobile deployment.[62] The MacPaw partnership combines LFMs designed for macOS assistance with MacPaw's Elix inference and Mnemos memory technologies; the same stack could later be offered to developers through MacPaw's Setapp marketplace.[60] In the LFM3 announcement, a MacPaw engineer said that on 65 everyday agent tasks in real macOS environments, LFM3-1.7B reached 64% task success against 48% for Qwen3.5-2B while using 58% fewer tokens per task.[78] With Insilico Medicine, Liquid AI trained LFM2-2.6B-MMAI on about 120 billion tokens of pharmaceutical data across more than 200 tasks using Insilico's MMAI Gym; the companies said it outperformed the more than ten-times-larger TxGemma-27B on 13 of 22 property-prediction tasks.[61]

## What other research has Liquid AI published?

Besides model releases, Liquid AI publishes research on sequence modelling and efficient architectures. In December 2025 it formally named its research team Liquid Labs, describing it as the group "driving fundamental breakthroughs in the science of building intelligent, personalized, and adaptive machines" and committing to publish technical reports, ablations, evaluations, and model weights.[81] Papers listed on its website in 2026 include "Preconditioned DeltaNet: Curvature-aware Sequence Modeling for Linear Recurrences" (April 2026), co-authored by Rus, and "In-Place Tokenizer Expansion for Pre-trained LLMs" (July 2026), co-authored by Amini and Lechner, which describes the procedure the company used to double the vocabulary of later LFM2.5 models.[75][77][49] In September 2026 it released LongevityBench, a benchmark for evaluating language models on aging-biology data such as clinical records, DNA methylation, transcriptomic, and proteomic profiles, together with fine-tuned LFM2-1.2B-Longevity and LFM2-2.6B-Longevity models and an accompanying paper published by Cell Press.[68]

## How does Liquid AI compare with related approaches?

The table below compares Liquid AI with another company commercializing an alternative to attention-only transformers and with several transformer-based model developers.

| Company | Founded | Architecture | Primary focus |
|---|---|---|---|
| Liquid AI | 2023 | Hybrid: gated short convolutions with a minority of grouped-query attention (LFM2 family); convolutions plus sliding-window and global attention (LFM3 family) | On-device and edge foundation models |
| [Cartesia](https://aiwiki.ai/wiki/cartesia) | 2023 | State space models | Real-time voice models |
| [Mistral AI](https://aiwiki.ai/wiki/mistral_ai) | 2023 | Transformer + Mixture-of-Experts | Open and commercial transformer LLMs |
| [DeepSeek](https://aiwiki.ai/wiki/deepseek) | 2023 | Transformer + MoE with [multi-head latent attention](https://aiwiki.ai/wiki/multi-head_latent_attention) | Open reasoning and coding LLMs |
| [Anthropic](https://aiwiki.ai/wiki/anthropic) | 2021 | Transformer | Frontier safety-focused models |
| [Google DeepMind](https://aiwiki.ai/wiki/google_deepmind) | 2010 (merged with Google Brain in 2023) | Transformer | Multimodal [frontier models](https://aiwiki.ai/wiki/frontier_models) |

[Cartesia](https://aiwiki.ai/wiki/cartesia), whose co-founders include [Albert Gu](https://aiwiki.ai/wiki/albert_gu), Karan Goel, Arjun Desai, and Brandon Yang,[72] also commercializes an alternative to attention-only transformers. Cartesia builds on state space model research, while Liquid AI's LFM2-family models are hybrids of gated short convolutions and a small number of attention blocks,[19] and its LFM3 models add sliding-window attention as a third operator.[78] The LFM2 technical report's architecture search included state-space blocks such as S4 and [Mamba](https://aiwiki.ai/wiki/mamba) and linear-attention variants; the report concluded that in the on-device regime most of the benefits attributed to recent hybrid state-space and linear-attention blocks can be captured by their short convolution submodules plus a small number of global attention layers.[19]

Open projects such as [RWKV](https://aiwiki.ai/wiki/rwkv) have also developed recurrent alternatives to attention. Liquid AI's own approach rests on its LIV-based design space, the STAR search method that lets it tune architectures to a hardware target, and a commercial focus on edge deployment.[37][5]

Liquid AI's business model also differs from labs that sell cloud access to large models. It combines open-weight releases that are free for smaller companies, commercial licences for companies above the $10 million revenue threshold, a paid API, and enterprise agreements such as the Shopify and Mercedes-Benz deals.[4][55][8][11]

## What are the strengths of Liquid AI's approach?

The most consistent claim in Liquid AI's releases is strong benchmark performance per parameter and per byte of memory: for example, the company reported that LFM2-1.2B performed competitively with Qwen3-1.7B, a model with a 47% bigger parameter count, and that LFM2 decoded and prefilled 2x faster than Qwen3 on CPU.[5] Most of these evaluations are run by the company itself; in 2026 Liquid AI moved toward published, reproducible device measurements with Pipette, built with Artificial Analysis.[57] The company also has commercial agreements with a chipmaker (AMD), a carmaker (Mercedes-Benz), and a commerce platform (Shopify), and Shopify runs an LFM in production for search.[10][11][8] The founders' peer-reviewed LTC and CfC papers give the company a published academic basis for its architecture work,[1][2] and the STAR architecture search lets it tune architectures to specific hardware targets.[16][37][5]

A second strength is the developer tooling. LEAP, Apollo, the open-weight Hugging Face releases, and builds for llama.cpp, MLX, ONNX, vLLM, and SGLang contrast with the closed initial LFM launch.[25][5][4] The LFM2 technical report discloses pretraining token counts, post-training pipelines, and architecture-search procedures.[19]

Outside researchers have pointed to efficiency as the appeal of the approach. Peter Bentley, a computer scientist at University College London who studies biologically inspired computing, told BBC Science Focus that AI is dominated by "massive power-hungry models" and that fewer neurons mean smaller models, less compute, and less energy.[73]

## What are the limitations and criticisms of Liquid AI?

The original LFM-1B, LFM-3B, and LFM-40B weights were not open, and the company did not publish their architecture.[15][35] WIRED noted in 2024 that the founders' liquid networks are especially suited to tasks involving temporal data, that making the technology work with other data types requires custom code, and that persuading big companies to base important projects on a completely new AI design would be another challenge.[33] Science News reported that the approach will not necessarily use less energy than mainstream AI, quoting Harvard computational neuroscientist Kanaka Rajan as saying that its constant adaptation makes it "computationally intensive."[34]

The LFM Open License's revenue threshold also limits the commercial reach of the open weights compared with pure Apache 2.0 releases: companies with more than $10 million in annual revenue must obtain a separate commercial licence to use them.[7][4]

Many of the benchmark advantages are measured by Liquid AI against open baselines of similar size. Hyena Edge, for example, was compared with a parameter-matched GQA-Transformer++ baseline that Liquid AI trained on the same 100 billion tokens,[23] and the LFM2 launch used the company's internal evaluation suite, which changed some details relative to EleutherAI's lm-evaluation-harness and evaluated Qwen3 in non-reasoning mode only.[5] Liquid AI's own figures show its larger models trailing bigger competitors on knowledge-heavy and coding benchmarks.[52] The LFM3 launch results were likewise produced by Liquid AI, which said larger models still led on agentic coding and knowledge.[78]

The architecture has been tested only at modest scale. The largest model Liquid AI has released was the 40.3-billion-parameter LFM-40B of 2024, and its largest current model is LFM2.5-24B-A2B, a sparse mixture of experts with about 2 billion active parameters.[35][52] The company reported that quality improved log-linearly from 350 million to 24 billion total parameters in the LFM2 family,[45] but it has not released a model with hundreds of billions of parameters. The first LFM3 release, in October 2026, covered only models of about 2 billion parameters.[78]

## What has happened with Liquid AI from 2023 to 2026?

| Date | Event |
|---|---|
| 2023 | Founded by Hasani, Lechner, Amini, and Rus[3][27] |
| December 6, 2023 | Emerges from stealth with a two-stage seed round of about $37.5 million; TechCrunch reports a $303 million post-money valuation for the tranche[27][21] |
| January 2024 | Capgemini collaboration announced at CES[9] |
| September 30, 2024 | First-generation LFM-1B, LFM-3B, and LFM-40B released in preview[35][14] |
| November 2024 | STAR architecture-search paper published[37][16] |
| December 13, 2024 | $250 million Series A led by AMD[6][12] |
| January 20, 2025 | LFM-7B released[36] |
| April 2025 | Hyena Edge demonstrated on a Samsung Galaxy S24 Ultra[22][23] |
| June 2025 | Sharp and G42 partnerships announced[62][63] |
| July 2025 | Open-weight LFM2 released (July 10); LEAP and Liquid Apollo launched (July 15)[5][25] |
| August 2025 | LFM2-VL released; LEAP support for AMD Ryzen; Alef Education collaboration[7][10][64] |
| September 2025 | LFM2-2.6B, Liquid Nanos, and the Brilliant Labs agreement[38][42][65] |
| October 2025 | LFM2-Audio, LFM2-8B-A1B, LFM2-VL-3B, and LFM2-ColBERT-350M released[40][39][74][41] |
| November 2025 | Shopify partnership (November 13); LFM2 technical report on arXiv (November 28); Hack the Edge with AMD[8][19][67] |
| December 4, 2025 | Research team formally named Liquid Labs[81] |
| January 2026 | LFM2.5 family introduced[43] |
| February 2026 | LFM2-24B-A2B early checkpoint released[45] |
| March 2026 | Insilico Medicine partnership and LFM2-2.6B-MMAI[61] |
| April 23, 2026 | Mercedes-Benz partnership announced[11] |
| May-August 2026 | LFM2.5-8B-A1B, LFM2.5-230M, LFM2.5-2.6B, LFM2.5-VL-3B, and DSpark drafts; MacPaw partnership (August 5); Pipette (August 24)[47][48][49][50][51][60][57] |
| September 2026 | Liquid Context on Snapdragon announced at Snapdragon Summit; LongevityBench released[58][68] |
| October 2026 | d1 decision models (October 5-15); LFM2.5-24B-A2B (October 13); RIPL inference engine (October 22); LFM3 family of six models (October 27)[53][54][55][52][56][78] |

Three threads run through that timeline. The first is a shift from closed-weight cloud models in 2024 toward openly licensed on-device models in 2025 and 2026, under the LFM Open License and with deployment tools such as LEAP.[15][5][25] The second is the relationship with AMD, which went from leading the Series A to technical work across Instinct data-center GPUs and Ryzen client processors, plus a jointly hosted hackathon.[6][10][67] The third is a growing list of customers and partners in latency-sensitive industries: commerce with Shopify, automotive with Mercedes-Benz, consumer devices with Sharp, Brilliant Labs, and MacPaw, and drug discovery with Insilico Medicine.[8][11][62][65][60][61]

As of October 2026 Liquid AI had not disclosed its revenue or a valuation more recent than its December 2024 Series A.[71]

## See also

- [List of neolabs](https://aiwiki.ai/wiki/list_of_neolabs)
- [Neolab](https://aiwiki.ai/wiki/neolab)
- [Foundation models](https://aiwiki.ai/wiki/foundation_models)
- [Language Model](https://aiwiki.ai/wiki/language_model)
- [LFM3](https://aiwiki.ai/wiki/lfm3)
- [Small language model](https://aiwiki.ai/wiki/small_language_model)
- [On-device AI](https://aiwiki.ai/wiki/on_device_ai)
- [Edge AI](https://aiwiki.ai/wiki/edge_ai)
- [Massachusetts Institute of Technology](https://aiwiki.ai/wiki/mit)
- [Robot foundation model](https://aiwiki.ai/wiki/robot_foundation_model)
- [Cartesia](https://aiwiki.ai/wiki/cartesia)
- [Hyena](https://aiwiki.ai/wiki/hyena)
- [Mamba](https://aiwiki.ai/wiki/mamba)
- [State space model](https://aiwiki.ai/wiki/state_space_model)
- [RWKV](https://aiwiki.ai/wiki/rwkv)
- [Recurrent neural network](https://aiwiki.ai/wiki/recurrent_neural_network)
- [Transformer](https://aiwiki.ai/wiki/transformer)
- [Hugging Face](https://aiwiki.ai/wiki/hugging_face)

## References

1. Hasani, R., Lechner, M., Amini, A., Rus, D., and Grosu, R. (2021). "Liquid Time-constant Networks." *Proceedings of the AAAI Conference on Artificial Intelligence* 35(9): 7657-7666. https://ojs.aaai.org/index.php/AAAI/article/view/16936
2. Hasani, R., Lechner, M., Amini, A., Liebenwein, L., Ray, A., Tschaikowski, M., Teschl, G., and Rus, D. (2022). "Closed-form continuous-time neural networks." *Nature Machine Intelligence* 4(11): 992-1003. https://www.nature.com/articles/s42256-022-00556-7
3. Liquid AI. "About | Liquid AI." Company page. https://www.liquid.ai/company/about
4. Liquid AI. "Liquid Foundation Models." Models page. https://www.liquid.ai/models
5. Liquid AI. "Introducing LFM2: The Fastest On-Device Foundation Models on the Market" (July 10, 2025). https://www.liquid.ai/blog/liquid-foundation-models-v2-our-second-series-of-generative-ai-models
6. Liquid AI. "We raised $250M to scale capable and efficient general-purpose AI" (December 13, 2024). https://www.liquid.ai/blog/we-raised-250m-to-scale-capable-and-efficient-general-purpose-ai
7. Liquid AI. "LFM2-VL: Efficient Vision-Language Models" (August 12, 2025). https://www.liquid.ai/blog/lfm2-vl-efficient-vision-language-models
8. Liquid AI. "Liquid AI Announces Multi-Year Partnership with Shopify to Bring Sub-20ms Foundation Models to Core Commerce Experiences" (November 13, 2025). https://www.liquid.ai/blog/liquid-ai-announces-multi-year-partnership-with-shopify-to-bring-sub-20ms-foundation-models-to-core-commerce-experiences
9. Liquid AI. "Liquid AI announces collaboration with Capgemini to build next-generation AI solutions for enterprises" (January 12, 2024). https://www.liquid.ai/blog/liquid-ai-announces-collaboration-with-capgemini-to-build-next-generation-ai-solutions-for-enterprises
10. Liquid AI. "Liquid's Edge AI Platform, LEAP, Expands Support to Laptops with Best-in-Class Performance on AMD Ryzen and Ryzen AI Processors" (August 18, 2025). https://www.liquid.ai/press/liquids-edge-ai-platform-leap-expands-support-to-laptops-with-best-in-class-performance-on-amd-ryzen-tm-and-ryzen-ai-tm-processors
11. Liquid AI. "Liquid AI and Mercedes-Benz partner to scale embedded in-car intelligence" (April 23, 2026). https://www.liquid.ai/press/liquid-ai-and-mercedes-benz-partner-to-scale-embedded-in-car-intelligence
12. TechCrunch. "Liquid AI just raised $250M to develop a more efficient type of AI model" (December 13, 2024). https://techcrunch.com/2024/12/13/liquid-ai-just-raised-250m-to-develop-a-more-efficient-type-of-ai-model/
13. Bloomberg. "Liquid AI Raising $250 Million to Build AI Inspired by Worm Brains" (December 13, 2024). https://www.bloomberg.com/news/articles/2024-12-13/liquid-ai-raising-250-million-to-build-ai-inspired-by-worm-brains
14. SiliconANGLE. "Liquid AI debuts new LFM-based models that seem to outperform most traditional large language models" (September 30, 2024). https://siliconangle.com/2024/09/30/liquid-ai-debuts-new-lfm-based-models-seem-outperform-traditional-llms/
15. VentureBeat. "MIT spinoff Liquid debuts non-transformer AI models and they're already state-of-the-art" (September 30, 2024). https://venturebeat.com/ai/mit-spinoff-liquid-debuts-non-transformer-ai-models-and-theyre-already-state-of-the-art
16. VentureBeat. "Liquid AI's new STAR model architecture outshines Transformer efficiency" (December 2, 2024). https://venturebeat.com/ai/liquid-ais-new-star-model-architecture-outshines-transformer-efficiency
17. MIT News. "Liquid machine-learning system adapts to changing conditions" (January 28, 2021). https://news.mit.edu/2021/machine-learning-adapts-0128
18. MarkTechPost. "Liquid AI Open-Sources LFM2: A New Generation of Edge LLMs" (July 13, 2025). https://www.marktechpost.com/2025/07/13/liquid-ai-open-sources-lfm2-a-new-generation-of-edge-llms/
19. Liquid AI. "LFM2 Technical Report." arXiv:2511.23404 (November 2025). https://arxiv.org/abs/2511.23404
20. Digital Commerce 360. "Shopify partners with Liquid AI on product search, recommendations" (November 13, 2025). https://www.digitalcommerce360.com/2025/11/13/shopify-liquid-ai-product-search-recommendations/
21. SiliconANGLE. "Liquid AI raises $37.6M to build 'liquid' neural networks" (December 6, 2023). https://siliconangle.com/2023/12/06/liquid-ai-raises-37-6m-build-liquid-neural-networks/
22. VentureBeat. "Liquid AI is revolutionizing LLMs to work on edge devices like smartphones with new 'Hyena Edge' model" (April 25, 2025). https://venturebeat.com/ai/liquid-ai-is-revolutionizing-llms-to-work-on-edge-devices-like-smartphones-with-new-hyena-edge-model
23. Liquid AI. "Convolutional Multi-Hybrids for Edge Devices" (April 24, 2025). Research page. https://www.liquid.ai/research/convolutional-multi-hybrids-for-edge-devices
24. MarkTechPost. "Liquid AI Introduces STAR: An AI Framework for the Automated Evolution of Tailored Architectures" (December 3, 2024). https://www.marktechpost.com/2024/12/03/liquid-ai-introduces-star-an-ai-framework-for-the-automated-evolution-of-tailored-architectures/
25. Liquid AI. "Liquid AI Launches LEAP and Liquid Apollo: The Easiest Way to Build with On-Device AI" (July 15, 2025). https://www.liquid.ai/blog/liquid-ai-launches-leap-and-apollo-bringing-edge-ai-to-every-developer
26. U.S. Securities and Exchange Commission. "Form D: Notice of Exempt Offering of Securities, Liquid AI, Inc." (filed December 18, 2023). https://www.sec.gov/Archives/edgar/data/2004369/000156761923007868/0001567619-23-007868-index.htm
27. Kyle Wiggers, TechCrunch. "Liquid AI, a new MIT spinoff, wants to build an entirely new type of AI" (December 6, 2023). https://techcrunch.com/2023/12/06/liquid-ai-a-new-mit-spinoff-wants-to-build-an-entirely-new-type-of-ai/
28. Chris McKay, Maginative. "AMD Leads $250M Investment in Liquid AI, Betting on Worm-Inspired AI Architecture" (December 13, 2024). https://www.maginative.com/article/amd-leads-250m-investment-in-liquid-ai-betting-on-worm-inspired-ai-architecture/
29. Mathias Lechner. "Mathias Lechner - Personal Page." https://mlech26l.github.io/
30. Alexander Amini. Personal website and biography. https://www.mit.edu/~amini/
31. MIT News. "Daniela Rus named CSAIL director" (May 22, 2012). https://news.mit.edu/2012/rus-named-csail-director-0522
32. Lechner, M., Hasani, R., Amini, A., Henzinger, T. A., Rus, D., and Grosu, R. (2020). "Neural circuit policies enabling auditable autonomy." *Nature Machine Intelligence* 2: 642-652. https://www.nature.com/articles/s42256-020-00237-3
33. Will Knight, WIRED. "Liquid AI Is Redesigning the Neural Network" (October 23, 2024). https://www.wired.com/story/liquid-ai-redesigning-neural-network/
34. Kathryn Hulick, Science News. "More brainlike computers could change AI for the better" (February 26, 2025). https://www.sciencenews.org/article/brainlike-computers-ai-improvement
35. Liquid AI. "Liquid Foundation Models: Our First Series of Generative AI Models" (September 30, 2024). https://www.liquid.ai/blog/liquid-foundation-models-our-first-series-of-generative-ai-models
36. Liquid AI. "Introducing LFM-7B: Setting New Standards for Efficient Language Models" (January 20, 2025). https://www.liquid.ai/blog/introducing-lfm-7b-setting-new-standards-for-efficient-language-models
37. Thomas, A. W., Parnichkun, R., Amini, A., Massaroli, S., and Poli, M. (2024). "STAR: Synthesis of Tailored Architectures." arXiv:2411.17800. https://arxiv.org/abs/2411.17800
38. Liquid AI. "Introducing LFM2-2.6B: Redefining Efficiency in Language Models" (September 23, 2025). https://www.liquid.ai/blog/introducing-lfm2-2-6b-redefining-efficiency-in-language-models
39. Liquid AI. "LFM2-8B-A1B: An Efficient On-device Mixture-of-Experts" (October 7, 2025). https://www.liquid.ai/blog/lfm2-8b-a1b-an-efficient-on-device-mixture-of-experts
40. Liquid AI. "LFM2-Audio: An End-to-End Audio Foundation Model" (October 1, 2025). https://www.liquid.ai/blog/lfm2-audio-an-end-to-end-audio-foundation-model
41. Liquid AI. "LFM2-ColBERT-350M: One Model to Embed Them All" (October 28, 2025). https://www.liquid.ai/blog/lfm2-colbert-350m-one-model-to-embed-them-all
42. Liquid AI. "Introducing Liquid Nanos: frontier-grade performance on everyday devices" (September 25, 2025). https://www.liquid.ai/blog/introducing-liquid-nanos-frontier-grade-performance-on-everyday-devices
43. Liquid AI. "Introducing LFM2.5: The Next Generation of On-Device AI" (January 5, 2026). https://www.liquid.ai/blog/introducing-lfm2-5-the-next-generation-of-on-device-ai
44. Liquid AI. "LFM2.5-1.2B-Thinking: On-Device Reasoning Under 1GB" (January 20, 2026). https://www.liquid.ai/blog/lfm2-5-1-2b-thinking-on-device-reasoning-under-1gb
45. Liquid AI. "LFM2-24B-A2B: Scaling Up the LFM2 Architecture" (February 24, 2026). https://www.liquid.ai/blog/lfm2-24b-a2b
46. Liquid AI. "LFM2.5-350M: No Size Left Behind" (March 31, 2026). https://www.liquid.ai/blog/lfm2-5-350m-no-size-left-behind
47. Liquid AI. "LFM2.5-8B-A1B: An Even Better On-Device Mixture of Experts" (May 28, 2026). https://www.liquid.ai/blog/lfm2-5-8b-a1b
48. Liquid AI. "LFM2.5-230M: Built to Run Anywhere" (June 25, 2026). https://www.liquid.ai/blog/lfm2-5-230m
49. Liquid AI. "LFM2.5-2.6B: Deploy Agents Everywhere" (August 4, 2026). https://www.liquid.ai/blog/lfm2-5-2-6b
50. Liquid AI. "LFM2.5-VL-3B: A Better and Faster Vision-Language Model for the Edge" (August 12, 2026). https://www.liquid.ai/blog/lfm2-5-vl-3b
51. Liquid AI. "LFM2.5-DSpark: Up to 3.2x Faster Inference from H100 to MacBook" (August 20, 2026). https://www.liquid.ai/blog/lfm2.5-dspark
52. Liquid AI. "LFM2.5-24B-A2B: Built for Fast On-Device Agents" (October 13, 2026). https://www.liquid.ai/blog/lfm2-5-24b-a2b
53. Liquid AI. "Introducing d1: The most capable decision model, now with vision" (October 5, 2026). https://www.liquid.ai/blog/d1-decision-model
54. Liquid AI. "Open d1: Edge decision models for text, vision, and audio" (October 7, 2026). https://www.liquid.ai/blog/d1-open
55. Liquid AI. "d1-ultra: Introducing the World's Best Decision Model" (October 15, 2026). https://www.liquid.ai/blog/d1-ultra
56. Liquid AI. "RIPL: Real-time Intelligence When Every Millisecond Counts" (October 22, 2026). https://www.liquid.ai/blog/low-latency
57. Liquid AI. "Introducing Pipette: A benchmarking suite for on-device intelligence" (August 24, 2026). https://www.liquid.ai/blog/pipette-on-device-ai-benchmarking-by-liquid-ai
58. Liquid AI. "Liquid AI, in Collaboration with Qualcomm Technologies, Brings Personal AI Context to Devices Powered by Snapdragon" (September 23, 2026). https://www.liquid.ai/blog/liquid-context-qualcomm
59. audioXpress. "Liquid AI Announces Personal AI Context to Devices Powered by Snapdragon Processors" (September 24, 2026). https://audioxpress.com/news/liquid-ai-announces-personal-ai-context-to-devices-powered-by-snapdragon-processors
60. Liquid AI. "MacPaw Partners with Liquid AI to Bring On-Device AI to Millions of Mac Users" (August 5, 2026). https://www.liquid.ai/blog/macpaw-partners-liquid-ai-on-device-ai-mac-users
61. Liquid AI. "Liquid AI and Insilico Medicine Announce Strategic Partnership Delivering Lightweight Scientific Foundation Models for Drug Discovery" (March 3, 2026). https://www.liquid.ai/press/liquid-ai-insilico-medicine-partnership
62. Liquid AI. "Liquid AI to Power On-Device Intelligence in Sharp's Bestselling Mobile Phone" (June 15, 2025). https://www.liquid.ai/press/liquid-ai-to-power-on-device-intelligence-in-sharps-bestselling-mobile-phone
63. Liquid AI. "G42 and Liquid AI Partner to Deliver Private, Local and Efficient AI Solutions to Enterprises at Scale" (June 17, 2025). https://www.liquid.ai/press/g42-and-liquid-ai-partner-to-deliver-private-local-and-efficient-ai-solutions-to-enterprises-at-scale
64. Liquid AI. "Alef Education Collaborates with Liquid AI to Advance AI in Education Globally" (August 19, 2025). https://www.liquid.ai/press/alef-education-collaborates-with-liquid-ai-to-advance-ai-in-education-globally
65. Liquid AI. "Brilliant Labs Partners With Liquid AI to Bring Vision-Language Tech to Your Glasses" (September 15, 2025). https://www.liquid.ai/press/brilliant-labs-partners-with-liquid-ai-to-bring-vision-language-tech-to-your-glasses
66. Repairer Driven News. "Mercedes announces multi-year partnership with Liquid AI" (April 27, 2026). https://www.repairerdrivennews.com/2026/04/27/mercedes-announces-multi-year-partnership-with-liquid-ai/
67. AMD. "Hack the Edge: Developers unleash creativity at AMD x Liquid AI Hackathon" (November 18, 2025). https://www.amd.com/en/developer/resources/technical-articles/2025/hack-the-edge-amd-and-liquid-ai-hackathon-recap.html
68. Liquid AI. "Setting the Frontier of Aging Biology with Liquid Foundation Models" (September 17, 2026). https://www.liquid.ai/blog/longevitybench
69. Liquid AI. "Careers." https://www.liquid.ai/careers
70. YXS Capital. "Liquid AI" portfolio page. https://www.yxscapital.com/portfolio/liquid-ai
71. Liquid AI. "Company news." News page. https://www.liquid.ai/news/company-news
72. Cartesia. "Company." https://cartesia.ai/company
73. Tom Howarth, BBC Science Focus. "This tiny worm's brain could transform artificial intelligence. Here's how" (November 7, 2025). https://www.sciencefocus.com/news/worm-brain-liquid-ai
74. Liquid AI. "LFM2-VL-3B: A New Efficient Vision-Language for the Edge" (October 22, 2025). https://www.liquid.ai/blog/lfm2-vl-3b-a-new-efficient-vision-language-for-the-edge
75. Liquid AI. "News and Updates." News page. https://www.liquid.ai/news
76. Ramin Hasani. "Ramin Hasani's Official Website." https://www.raminhasani.com/
77. Liquid AI. "Research." Publications page listing paper titles, dates and authors. https://www.liquid.ai/news/research
78. Liquid AI. "Introducing LFM3: Every Modality for Specialized Intelligence" (October 27, 2026). https://www.liquid.ai/blog/lfm3
79. Liquid AI. "LFM3-1.7B" model card. Hugging Face, accessed October 29, 2026. https://huggingface.co/LiquidAI/LFM3-1.7B
80. MarkTechPost. "Liquid AI Releases LFM3: 6 Open-Weight 2B-Class Multimodal Models" (October 27, 2026). https://www.marktechpost.com/2026/10/27/liquid-ai-lfm3-open-weight-2b-multimodal-models/
81. Liquid AI. "Introducing Liquid Labs: Defining the Frontier of Intelligence through Innovation" (December 4, 2025). https://www.liquid.ai/blog/introducing-liquid-labs-defining-the-frontier-of-intelligence-through-innovation
82. Radical Numerics. "About Radical Numerics." Accessed October 29, 2026. https://www.radicalnumerics.ai/about
83. Lily Mae Lazarus, Fortune. "Exclusive: The researchers who built AI-generated DNA just raised $50 million to reinvent biology" (June 15, 2026). https://fortune.com/2026/06/15/exclusive-ai-dna-radical-numerics-eric-nguyen-biology-biodefense-drug-discovery/
84. Liquid AI. "LiquidAI/d1-2B" model card. Hugging Face, accessed October 29, 2026. https://huggingface.co/LiquidAI/d1-2B