Als nächstesNächster Leitfaden
NVIDIA Cosmos World Foundation-Modelle
Unternehmen
Leitfaden für Unternehmen
Liquid Foundation Models (LFMs) are a family of AI models from Liquid AI.
Hören Sie sich diesen Leitfaden anKI-erzählt · ~8 min
LFM2 combines short convolutions with attention to run language tasks on devices with limited memory, including phones and laptops.
Liquid AI is an American AI company spun out of the Computer Science and Artificial Intelligence Laboratory (CSAIL) at MIT. TechCrunch names robotics researcher Daniela Rus as a co-founder, and the company grew out of liquid neural network research by Rus and her colleagues. The company releases and sells Liquid Foundation Models, or LFMs: general-purpose generative models built to be small and fast enough to run on edge devices, meaning hardware at the point of use such as smartphones, laptops, cars and robots, rather than on large server clusters. A foundation model is a model trained on broad data that can then be adapted to many tasks.
The company's name comes from its research roots. In a 2020 paper accepted to AAAI-21, Hasani, Lechner, Amini, Rus and Radu Grosu introduced liquid time-constant networks. These are recurrent neural networks, models that process a sequence one step at a time while carrying a hidden state, whose neurons follow simple differential equations. The twist is that each neuron's time constant, which controls how quickly it reacts to new input, is itself changed by a learned gate that depends on the input. The authors showed the networks stay stable and bounded and reported better time-series prediction than classical and modern recurrent networks. TechCrunch described the idea as inspired by the brain of a roundworm and noted that such networks can be much smaller than conventional models.
The commercial LFMs are not literal copies of those 2020 networks. They borrow the broader idea of input-dependent operators. Liquid AI's July 2025 LFM2 announcement frames its design space around what it calls linear input-varying (LIV) operators: linear operations whose weights are generated on the fly from the input they act on. Under that view, convolutions, recurrences and attention are all special cases, so an automated search can mix them. The company built a neural architecture search engine called STAR to explore that space, and for LFM2 it scored candidates by measured peak memory and prefill and decode speed on Qualcomm Snapdragon phone CPUs instead of relying on theoretical proxies such as cache size.
The search produced a hybrid backbone. Each LFM2 model in the first release has 16 blocks: 10 double-gated short convolution blocks and 6 grouped-query attention blocks, with a SwiGLU feed-forward layer and RMSNorm normalization in each block. A short convolution blends each token only with a few neighbors, so its cost grows linearly with text length and it needs no growing memory cache. Grouped-query attention, a cheaper variant of standard attention in which several query heads share one set of keys and values, handles the long-range lookups that convolutions cannot. Keeping attention to a minority of blocks is what shrinks memory use and speeds up generation on CPUs. The blog says the choice of short convolutions follows from the target hardware, embedded system-on-chip CPUs, and from how well existing kernel libraries run them.
The milestones are recent. Liquid AI released its first LFM series on 30 September 2024: LFM-1B (about 1.3 billion parameters), LFM-3B (about 3.1 billion) and LFM-40B, a mixture-of-experts model with 40.3 billion total parameters and roughly 12 billion active per token, all with a 32,000-token context window. That generation is now listed as deprecated. In December 2024 the company raised a $250 million Series A led by AMD, which TechCrunch reported at a valuation above $2 billion citing Bloomberg, with plans to tune models for AMD GPUs, CPUs and accelerators. LFM2 followed on 10 July 2025 with open-weight 350M, 700M and 1.2B checkpoints, and the LFM2 technical report posted to arXiv on 28 November 2025 extended the family to a 2.6B dense model and an 8.3B mixture-of-experts model with 1.5B active parameters. Newer LFM2.5 checkpoints, such as LFM2.5-1.2B-Instruct, keep the same 16-block layout but are pretrained on 28 trillion tokens and add multi-stage reinforcement learning.
Compared with the obvious alternatives, small transformer models such as Qwen3, Llama 3.2 and Gemma 3, the trade is speed and memory against breadth. Liquid AI's own table shows LFM2-1.2B scoring 55.23 on MMLU, a general-knowledge test, against 59.11 for the larger Qwen3-1.7B and 46.6 for Llama-3.2-1B-Instruct, while beating all three on IFEval, an instruction-following test, with 74.89. The company reports LFM2-700M decoding faster than Qwen3-0.6B in both ExecuTorch and llama.cpp despite being 16% larger. Compared with pure state-space models such as Mamba, which replace attention entirely with a recurrence, LFM2 keeps a few attention blocks for precise recall over long inputs.
The limitations matter. Nearly every benchmark and speed figure comes from Liquid AI itself, and its blog says it modified some evaluation procedures relative to the standard EleutherAI evaluation harness, so results are not directly comparable with other leaderboards. Qwen3 was tested with its reasoning mode turned off. Models of one to three billion parameters still hold far less world knowledge than frontier models and can be confidently wrong, so they suit narrow tasks such as extraction, classification and tool calling better than open-ended factual questions. Speed claims also depend on quantization: the published phone and laptop tests used 4-bit weights, which trade some accuracy for size.
In practice the requirements are modest. The models ship as open weights on Hugging Face with deployment packages for ExecuTorch, llama.cpp and vLLM, so a developer can run a 1.2B model on an ordinary laptop or phone without a GPU. The license is based on Apache 2.0 but adds a revenue threshold: academic and research use is free, commercial use is free for companies under $10 million in revenue, and larger companies must contact Liquid AI for a commercial license. Organizations that want custom models or enterprise support deal with the company directly.
04Ausgearbeitetes Beispiel
Install llama.cpp, one of the runtimes Liquid AI supports, on a laptop with no discrete GPU.
Download a 4-bit GGUF build of LFM2-1.2B and of a similar-size transformer such as Llama-3.2-1B-Instruct from Hugging Face.
Run llama.cpp's benchmark tool on each model with the same prompt length, for example 512 tokens in and 128 tokens out, and the same thread count.
Record prefill speed (tokens per second while reading the prompt) and decode speed (tokens per second while generating) for both models.
Repeat with a longer prompt, such as 4,096 tokens, to see how each model's speed and memory use change as context grows.
Ask both models the same ten factual and instruction-following questions and mark which answers are correct.
Was es zeigt
A small table of prefill speed, decode speed and accuracy for both models on one machine. Comparing it with Liquid AI's published 2x CPU claim shows how much the result depends on chip, runtime and quantization.
Running LFM2-1.2B fully offline on a laptop CPU through llama.cpp with Q4_0 quantization, one of the two runtimes Liquid AI used for its published speed tests on an AMD Ryzen HX370.
Shipping a private on-phone assistant with ExecuTorch, the runtime Liquid AI benchmarked on a Samsung Galaxy S24 Ultra, so prompts and replies never leave the device.
Using LFM2-Audio, described in the November 2025 LFM2 technical report, for real-time speech-to-speech conversation with separate pathways for audio input and audio output.
Building multilingual document search with LFM2-ColBERT, the report's low-latency encoder for queries and documents, on hardware without a dedicated GPU.
Feinabstimmung eines offenen LFM2-Kontrollpunkts für eine begrenzte Aufgabe wie die Datenextraktion, die mit der LFM2-Lizenz von Unternehmen mit einem Jahresumsatz von weniger als 10 Millionen US-Dollar kostenlos kommerziell genutzt werden kann.
Die Richtung, die Liquid AI signalisiert hat, sind effiziente Modelle für mehr Modalitäten und nicht immer größere dichte Netzwerke. Der technische Bericht vom November 2025 deckt bereits Vision-Sprach-, Audio- und Retrieval-Varianten sowie ein 8,3B-Expertenmix-Modell mit 1,5B aktiven Parametern ab, und die LFM2,5-Checkpoints erhöhen das Vortraining auf 28 Billionen Token mit Verstärkungslernen an der Spitze. Es bleiben offene Fragen. Unabhängige Bewertungen von hybriden Faltungs- und Aufmerksamkeitsdesigns sind im Vergleich zu denen für Transformatoren immer noch dürftig, daher ist nicht geklärt, wie gut der Geschwindigkeitsvorteil über Chips, Laufzeiten und Quantisierungsniveaus hinaus anhält, die über die vom Unternehmen getesteten hinausgehen. Es ist auch unklar, inwieweit kleine Modelle die Wissenslücke zu Grenzsystemen schließen können oder ob ihre Hauptaufgabe darin besteht, als schnelle lokale Agenten zu fungieren, die schwierigere Anforderungen an größere Cloud-Modelle weiterleiten. Die im Dezember 2024 angekündigte Partnerschaft mit AMD weist auf hardwarespezifisches Tuning hin, das die Architektursuche anhand realer Gerätemessungen, den STAR-Ansatz, zu einem wahrscheinlichen Bereich weiterer Arbeit macht.
Kostenloser Newsletter
Drei verifizierte KI-Geschichten jeden Werktagmorgen, in einfachem Englisch geschrieben. Für immer kostenlos, keine Werbung.
Eine E‑Mail jeden Werktag. Abmelden mit einem Klick. Wir verkaufen oder teilen Ihre Adresse nie.
Testen Sie sich selbst
Sofortiges Feedback zu jeder Antwort und ein teilbares Zertifikat mit verifizierbarer ID, sobald Sie einen Kurs bestanden haben.
Unterstützen Sie kostenlose KI-Bildung. AI Understanding ist eine 501(c)(3) gemeinnützige Organisation – keine Werbung, keine Paywall, nie. Machen Sie eine Spende
Liquid Foundation Models (LFMs) are a family of AI models from Liquid AI. LFM2 combines short convolutions with attention to run language tasks on devices with limited memory, including phones and laptops.
Der Bericht erweitert LFM2 als Familie der Liquid Foundation Models und beschreibt seine Modellvarianten. Diese Frage betrifft die Benennung des Berichts, nicht die biologische Inspiration.
Der LFM2-Bericht vom November 2025 beschreibt ein kompaktes Hybrid-Backbone mit Gated Short Convolutions und einer kleinen Anzahl von Aufmerksamkeitsblöcken für gruppierte Abfragen; Dies ist eine modellspezifische Architektur und kein Anspruch auf konstanten Speicher für jedes LFM.
Der technische Bericht listet eine Mix-of-Experts-Variante mit 8,3 Milliarden Gesamtparametern und 1,5 Milliarden aktiven Parametern auf. Dies ist eine Konfigurationstatsache für das benannte Modell.
Der LFM2-Bericht listet eine Kontextlänge von 32 KB für die Familie auf. Weitere Konfigurationsdetails finden Sie in der jeweiligen Variante und im Bereitstellungspaket.
Der Bericht beschreibt LFM2 als für die Bereitstellung auf dem Gerät unter geräteseitigen Latenz- und Spitzenspeicherbeschränkungen konzipiert und bewertet CPUs der Smartphone- und Laptop-Klasse. Hierbei handelt es sich um ein Designziel und einen Benchmark-Bereich, nicht um einen Nachweis der Bereitstellung auf jedem aufgeführten Gerät oder um eine Datenschutzgarantie.
Lerne weiter
Weitere Leitfäden zu diesem Thema ausgewählt
Als nächstesNächster Leitfaden
NVIDIA Cosmos World Foundation-Modelle
Unternehmen