Introducing LFM3: Every Modality for Specialized Intelligence

LFM3 is a family of compact, open-weight models designed for specialized tasks and efficient deployment on phones, computers, and other devices. Its first release comprises six models of roughly 2 billion parameters, sharing a backbone and covering text, vision, audio, multimodal “omni” use, and decision-making. The 1.7B base model was trained on 22 trillion tokens, including a final phase of curated data and agentic traces that extended its context window to 256K tokens. Its hybrid architecture combines short convolutions, sliding-window attention, and global attention, selected through hardware testing across phones, laptops, CPUs, GPUs, and an NPU. The design aims to add capability while keeping long-context inference fast; at 8K tokens, the 1.7B model is reported to run about 25% faster than LFM2-1.2B.

The text model is tuned for agents, tool use, coding, and instruction following. In reported evaluations, it performs strongly on function calling, instruction adherence, and several browser and business-workflow tasks, sometimes surpassing much larger models, though larger systems remain ahead in areas such as agentic coding and knowledge. It is also reported to decode faster and use less memory than comparison models on-device. The vision model adds video input, stepwise reasoning, object localization, and spatial understanding; across 62 vision benchmarks, it scores well above similar-sized models and approaches some systems many times larger. The central proposition is that developers can choose a modality-specific model, fine-tune it for a domain, and run many specialized AI workloads locally rather than relying on one large general-purpose model.