Liquid AI, a distinguished innovator in the artificial intelligence sector, has recently announced the launch of its latest foundational model suite, LFM3. The inaugural release encompasses six models, each boasting around 2 billion parameters and spanning diverse domains such as text, vision, audio, full modality, and decision-making capabilities. All these models are built upon the LFM3-1.7B-Base framework and leverage a staggering 22 trillion tokens of data during their pre-training phase. The LFM3 series adopts a three-tier hybrid layer architecture that seamlessly integrates gated short convolutions, sliding window attention mechanisms, and global attention. Following an architectural overhaul, it can seamlessly adapt to 14 different device and backend configurations. In comparison to its forerunners, the LFM3 series exhibits a roughly 25% increase in processing speed for 8K contexts. Moreover, the 4-bit quantized text model is compact, occupying only about 1GB of space, thereby substantially reducing the hurdles for edge deployment. Each of the six initial models is tailored to specific scenarios and is user-friendly for fine-tuning, facilitating cost-effective vertical industrial applications. Certain models also excel in delivering efficient real-time inference in scenarios involving embodied intelligence. Presently, the open-source weights of the LFM3 series are accessible and have garnered immediate support from numerous companies and mainstream development framework platforms.