Introducing LFM3: Every Modality for Specialized Intelligence

Smaller, more capable multimodal models are now available for local fine-tuning and on-device deployment.

Liquid AI has unveiled LFM3, a new generation of Liquid Foundation Models designed for specialized intelligence. The LFM3 family is multimodal, integrating text, vision, audio, omni, and decision models onto a single shared backbone. The initial release features models in the 2B parameter range, built on a new hybrid architecture and trained on 22 trillion tokens. This new architecture enhances expressivity while maintaining efficiency. LFM3 models are designed for on-device deployment and can be fine-tuned on a single GPU, making them suitable for specialized tasks. The base and post-trained models are open-weight and available on Hugging Face, the Liquid AI API, and their Playground. The company highlights LFM3's improved capability per parameter, multimodal integration, edge-device suitability, and fine-tunability for specific applications.

The LFM3 architecture employs a novel "tri-hybrid" layer mix combining gated short convolutions, sliding-window attention, and global attention, optimized for agentic use cases requiring high capability per parameter and fast inference over long contexts. This design allows LFM3-1.7B to outperform previous, larger models while using fewer parameters and fitting into approximately 1GB at 4-bit precision, making it suitable for edge devices. The models' extensive training on 22T tokens, including a significant portion focused on agentic traces and tool calls, prepares them for sophisticated on-device AI applications.

The release of a multimodal family of models (text, vision, audio, omni, decision) from a single backbone signifies a move towards more integrated and specialized AI deployments on consumer hardware. Partners like AMD, Ambarella, and Qualcomm are already integrating LFM3 into their platforms, indicating a market trend towards powerful, customizable AI that can run locally, enhancing privacy and reducing latency for specialized intelligent agents and applications.