Liquid AI is a startup founded in 2023 by computer scientists from the Massachusetts Institute of Technology. In July 2025, the company released the second generation of its Liquid Foundation Models series (LFM2), designed to provide the fastest on-device foundation models using a new "liquid" architecture. Its training and inference efficiency makes it a strong competitor to large language models in the cloud, such as OpenAI's GPT series and Google's Gemini.
LFM2 initially launched dense checkpoints with 350M, 700M, and 1.2B parameters, using a hybrid architecture that favors gated short convolutions. Benchmarking results show that LFM2 outperforms competitors like Qwen3, Llama3.2, and Gemma3 in both quality and CPU throughput. For enterprises, this series enables real-time and privacy-protected AI applications on devices such as smartphones, laptops, and vehicles, without sacrificing capability or latency.
After the release of LFM2, Liquid AI further expanded its product line, adding task- and domain-specific variants, small video ingestion and analysis models, and a deployment stack called LEAP. The company now published a detailed 51-page technical report on arXiv, revealing the architecture search process, training data mixing, distillation targets, curriculum strategies, and post-training processes behind the model. This blueprint will provide other organizations with a reference to train their own small and efficient models tailored to their hardware and deployment constraints, starting from scratch.
LFM2 was designed with real-world enterprise needs in mind, such as latency budgets, memory limits, and thermal thresholds, ensuring stable performance across various devices. The report also highlights its optimized training pipeline, aimed at achieving more reliable instruction execution and tool usage behavior, making LFM2 models more practical in real-world applications.
Key Points:
🔍 Small models can run efficiently on device, providing real-time privacy-protected AI services.
💡 The blueprint released by Liquid AI allows other companies to refer to it and build their own small and efficient models.
🚀 LFM2 models enhance predictability and ease of use on various hardware through optimized design.
