According to a report from Electronic Enthusiasts (by Li Wanwan), the American startup Liquid AI has officially announced the launch of its next-generation Liquid Foundation Model (LFM2), which sets new records in speed, energy efficiency, and quality within the edge model category.The birth of LFM2 stems from a reconstruction of the underlying logic of AI. Unlike traditional models based on Transformers, LFM2 is built using structured, adaptive operators, inspired by a deep integration of dynamic systems theory, signal processing, and numerical linear algebra.This design provides the model with three core advantages: Firstly, training efficiency has improved by 300% compared to the previous generation, with throughput on CPUs increased by 200% over models like Qwen3 and Gemma, significantly reducing latency; Secondly, it excels in long context processing (such as analyzing 32k token documents) and in resource-constrained scenarios (like mobile devices and IoT devices); Thirdly, by optimizing memory usage (LFM-3B requires only 16GB of memory, far less than Meta Llama-3.2-3B’s 48GB), it achieves millisecond-level real-time inference and offline operation.“We build models from first principles, as rigorously as engineers design engines.” emphasized Ramin Hasani, co-founder and CEO of Liquid AI, at the launch event, “the goal of LFM2 is to turn every device into a local AI hub, unleashing the potential of generative AI in edge scenarios.”From empirical data, LFM2 significantly outperforms competing products of similar scale in key attributes of agent AI, such as instruction following and function calling, making it an ideal choice for local and edge use cases; in authoritative evaluations like MMLU (Multimodal Learning Understanding) and RULER (Long Context Reasoning), the LFM2-3B model outperforms previous 7B-13B models, while LFM2-40B achieves a balance of quality and cost with a mixture of experts (MoE) architecture; In terms of ecological compatibility, the model weights have been open-sourced to Hugging Face, supporting instant testing through Liquid Playground, and will soon be integrated into edge AI platforms and iOS native applications.The launch of LFM2 coincides with the eve of a global explosion in the edge AI market. It is predicted that edge AI technologies in consumer electronics, robotics, smart home appliances, and other fields could drive the total potential market size close to $1 trillion by 2035. Liquid AI has partnered with Fortune 500 companies to provide ultra-high-efficiency small multimodal foundation models and secure enterprise-level deployment solutions, helping clients achieve data sovereignty and privacy protection.“Moving large generative models from the cloud to the device is a necessary path for the democratization of AI.” Hasani pointed out, “the millisecond-level latency and offline capabilities of LFM2 are crucial for terminals with high real-time requirements, such as smartphones, cars, and satellites.” Currently, the company has established strategic partnerships in defense, aerospace, and cybersecurity, with its technology stack becoming the preferred solution for high-security scenarios.The release of LFM2 is just the first step in Liquid AI’s technological journey. It has been revealed that in the coming months, the company will successively launch a series of powerful models covering a wider range of modalities and scenarios. As enterprises shift from cloud-based LLMs to economical, fast, and private local intelligence, LFM2 may become a key driver of AI technology democratization.
Disclaimer: This article is originally from Electronic Enthusiasts and must be cited as above. For group discussions, please add WeChat elecfans999, for submission of news tips or interview requests, please email [email protected].
More Hot Articles to Read
-
Domestic DRAM leader initiates IPO! Valuation nearly 140 billion
-
1 second transformation! Zhi Hui Jun launches “Nezha” wheeled robot, humanoid robot industry chain poised for takeoff
-
Domestic GPU makes another breakthrough, rising against Nvidia + AMD
-
Market value nearly $4 trillion! Nvidia GB300 server officially shipped
-
Work stoppage and production halt! Former charging treasure champion “Roma Shi” faces bankruptcy rumors? Official response
Click to Follow and Star Us
Set us as a star to not miss any updates!