Liquid AI Releases LFM3 Model Family with Novel Triple-Hybrid Architecture
Artificial intelligence laboratory Liquid AI has released the LFM3 model family, featuring text, vision, audio, omni, and decision-making models built on a shared 1.7B base pre-trained on 22T tokens. The initial lineup includes LFM-1.7B, LFM3-VL-2B, d1-2B, LFM3-Audio, LFM3-Omni-2B, and d1-Omni-2B. This release marks a significant milestone in efficient AI development by delivering high performance across multiple modalities while using fewer parameters and a lower footprint. It enables developers to run and fine-tune specialized intelligence models more easily on resource-constrained hardware. The LFM-1.7B model features a 256K context window and outperforms the previous-generation LFM2.5-2.6B despite using 35% fewer parameters. All models in the LFM3 family are based on a novel triple-hybrid layer architecture and can be fine-tuned on a single GPU.
## BACKGROUND
Liquid AI is an MIT-founded company that builds Liquid Foundation Models (LFMs) using liquid neural network principles as an alternative to traditional attention-based architectures. These models are designed for efficient deployment on edge devices, on-premise hardware, and cloud environments with low latency.