Liquid AI has released the LFM3 family, consisting of six open-weight multimodal models. Each model contains approximately 2 billion parameters and is designed to operate and undergo fine-tuning on a single GPU. The lineup includes specialized versions for text, vision-language, audio, decision-making, and an omni model capable of processing text, images, video, and audio simultaneously.
The models are built on the LFM3-1.7B-Base foundation and utilize a tri-hybrid architecture. This architecture integrates gated short convolutions, sliding-window attention, and global attention, having been pre-trained on 22 trillion tokens. The text-focused LFM3-1.7B model supports a context window of 256K tokens, while the vision-language variant, LFM3-VL-2B, reportedly matches the performance of models six times its size.
Founded in 2023, Liquid AI has secured approximately $287.5 million in total funding. This includes a $250 million Series A round completed in December 2024. The company aims to support edge computing applications, such as robotics and industrial defect detection, by enabling local deployment and customization of AI models without reliance on cloud infrastructure.
