Build specialized multimodal models with LFM3

Why it matters

LiquidAI has announced LFM3, its newest generation of Liquid Foundation Models, positioning the family as a base for specialized, on-device AI rather than one large general-purpose system. The release places six models covering text, vision, audio, omni, and decision tasks on a single shared backbone, LFM3-1.7B-Base, all within the 2B parameter class. The company says the flagship text model, LFM3-1.7B, outperforms its predecessor LFM2.5-2.6B despite having fewer parameters, with the largest gains on agentic tasks such as browsing and business-app workflows. Architectural changes include a tri-hybrid layer mix combining gated convolutions with sliding-window and global attention, head-wise gated attention to ease quantization, partial rotary position embeddings, and sandwich normalization. The design targets edge deployment: the text model fits in roughly one gigabyte at low-bit precision, fine-tuning runs on a single GPU, and the context window extends far enough for on-device agents. The vision model adds video input, and the audio model supports streaming with fast response times.