Build specialized multimodal models with LFM3
huggingface ·
LiquidAI has introduced LFM3, a new generation of Liquid Foundation Models designed to serve as a shared backbone for specialized multimodal intelligence. This architecture consolidates text, vision, audio, omni-modal, and decision-making capabilities into six distinct models, all ranging around the 2B parameter scale. The underlying LFM3-1.7B-Base utilizes a tri-hybrid layer mix and hardware-in-the-loop search to achieve higher expressiveness with 35% fewer parameters than its predecessor, LFM2.5. These models are optimized for edge deployment, featuring low-latency inference, small memory footprints, and support for long context windows up to 256K tokens. Key releases include LFM3-1.7B for agentic tasks, LFM3-VL-2B for visual understanding, and LFM3-Omni-2B for integrated audio-visual interaction. Additionally, the d1 series offers fast, structured decision-making capabilities for applications requiring calibrated answers rather than generated text. All models are open-weight and can be fine-tuned on a single GPU to adapt to specific domains, terminology, or robotic control tasks.
huggingface