Liquid AI profile, products and governance

The company's product line is the LFM (Liquid Foundation Model) family, which by mid-2026 spanned more than 50 shipped models and over 3,100 variants, from 230-million-parameter encoders up to a 24-billion-parameter mixture-of-experts model, all distributed primarily through Hugging Face. On October 27, 2026, Liquid AI announced LFM3, a new generation that puts text, vision, audio, omni, and decision models on one shared backbone. The first LFM3 release is six models in the 2-billion-parameter range, trained from LFM3-1.7B-Base on 22 trillion tokens, with open weights on Hugging Face. The previous generation, LFM2.5, still covers text models (LFM2.5-350M, LFM2.5-1.2B-Instruct, LFM2.5-2.6B, LFM2.5-8B-A1B), vision-language models (LFM2.5-VL-450M, LFM2.5-VL-1.6B, LFM2.5-VL-3B), an audio model (LFM2.5-Audio-1.5B), and retrieval-focused nano models (LFM2.5-Embedding-350M, LFM2.5-ColBERT-350M). Alongside the models, Liquid AI ships LEAP, a software development kit for fine-tuning and deploying LFMs across llama.cpp, MLX, ONNX, CoreML, SGLang, and vLLM runtimes, and Apollo, its hosted access layer for teams that want the same models without managing on-device runtimes themselves.

Liquid AI has shipped a new model roughly every month since March 2026: LFM2.5-350M in March, the vision-language LFM2.5-VL-450M in April, the mixture-of-experts LFM2.5-8B-A1B in May, LFM2.5-2.6B on August 4, and the vision-language LFM2.5-VL-3B on August 12. October added the d1 decision models (d1 on October 5, an open d1 on October 7, d1-ultra on October 15), LFM2.5-24B-A2B for on-device agents on October 13, and the LFM3 family on October 27. According to Liquid AI, LFM3-1.7B outperforms LFM2.5-2.6B with 35% fewer parameters, the text model fits in about 1 GB at 4-bit precision with a 256K context window, and every LFM3 model can be fine-tuned on a single GPU. On the partnership side, Mercedes-Benz announced in April 2026 that it would embed LFMs directly into third- and fourth-generation MBUX infotainment systems for North American vehicles starting in the second half of 2026, and Alef Education said in January 2026 that it was deploying Liquid AI's on-device models across a platform used by more than 1.5 million students in 14,000 K-12 schools.

Liquid AI has raised $297M across two rounds from 22 disclosed investors. The most recent, a $250M Series A that closed December 13, 2024, was led by AMD Ventures with participation from Andreessen Horowitz, Capgemini Ventures, and Breyer Capital, and valued the company at roughly $2.35B, making it a unicorn a little over a year after founding. As of August 2026, Liquid AI has not disclosed a newer round.

Revenue comes from the commercial side of the LFM Open License (LFM1.0): the models themselves, built on an Apache 2.0 base, are free to download, fine-tune, and deploy commercially for any company under $10M in annual revenue, and Liquid AI charges companies above that threshold for a commercial license, plus custom architecture work, OEM integration (as with Mercedes-Benz), and on-premises deployment support with SLAs. Liquid AI has not published official revenue figures; third-party estimator Getlatka put 2025 ARR at approximately $13.2M, a figure Liquid AI has not confirmed.

CEO Ramin Hasani and CTO Mathias Lechner both came out of Daniela Rus's lab at MIT CSAIL, where Hasani's doctoral research on liquid time-constant networks originated; Alexander Amini, the company's Chief Science Officer, was a CSAIL research scientist on the same liquid-networks work. Rus remains a co-founder and director while continuing to lead CSAIL. Third-party trackers put Liquid AI's headcount at roughly 100 to 150 people as of mid-2026 (Tracxn cites 130 as of June 30, 2026), spread across its Cambridge headquarters and a smaller Boston-area presence.

Liquid AI describes its mission as building efficient, general-purpose AI at every scale, with products that are compute- and cost-optimized enough to bring intelligence to any device. The company frames its research philosophy as white-box explainability, favoring architectures whose behavior can be inspected over black-box shortcuts, and publishes a model card for every LFM release on Hugging Face with benchmark tables, intended use, and stated limitations. Liquid AI has not published a formal responsible-scaling policy or named third-party red-teaming partners, in contrast to frontier labs like Anthropic or OpenAI.

Liquid AI competes for the on-device and edge tier of the model market against Google's Gemma family, Meta's Llama edge and mobile variants, and Microsoft's Phi series, rather than against cloud-first frontier labs. Its differentiator is architecture: LFMs mix convolution and attention rather than using a pure Transformer stack, which Liquid AI says lets a 3B-parameter vision-language model like LFM2.5-VL-3B run in about 3GB of memory at 228 tokens per second on an Apple M5 Max. The tradeoff is capability ceiling: Liquid AI has no flagship model built to compete with GPT-5, Claude, or Gemini on frontier reasoning benchmarks, and its own LFM2.5-VL-3B model card recommends against using it for long-context or reasoning-heavy tasks. The LFM lineup has passed 38 million downloads on Hugging Face, a proxy for developer adoption relative to other small-model providers.

Because LFMs run locally rather than through a hosted API by default, inference does not require sending user data to Liquid AI's servers, which the company markets as a privacy-by-architecture advantage over cloud-only competitors. Liquid AI has not published SOC 2, ISO 27001, or GDPR compliance certifications, and has not disclosed an EU AI Act classification as of August 2026.

The Mercedes-Benz MBUX rollout, scheduled for the second half of 2026, is Liquid AI's first large-scale automotive production deployment and the clearest signal of whether its edge-AI thesis can scale beyond developer downloads into shipped consumer hardware. The company's release cadence, a new model roughly every month through 2026, including the October LFM3 generation, suggests it intends to keep competing on frontier-efficiency (best capability per gigabyte) rather than frontier-scale.