Introducing LFM3: Every Modality for Specialized Intelligence (www.liquid.ai)

Today marks the launch of LFM3, a new generation of Liquid Foundation Models (LFMs) designed for specialized intelligence and featuring multimodal capabilities. This innovative model family combines text, vision, audio, and decision-making models into a single architecture, with the first release offering models around 2 billion parameters. LFM3 enhances expressiveness and efficiency through a hybrid architecture, pre-trained on 22 trillion tokens, making it ideal for localized fine-tuning on various devices. Notably, LFM3 achieves a significant performance upgrade over its predecessor, LFM2, particularly in long-context tasks, proving to be about 25% faster at processing 8K tokens. The implications for the AI/ML community are profound, as LFM3 boasts a tri-hybrid layer mix—including gated short convolutions, sliding-window attention, and global attention—all optimized for agentic tasks. The model shows competitive performance against larger counterparts, excelling in instruction following and function calling, and allowing for customizable deployments that can run on both consumer devices and cloud infrastructures. With features such as a 256K-token context window and the ability to handle complex tasks in real-time, LFM3 is poised to redefine how specialized intelligence applications are developed and implemented across various domains.