Build specialized multimodal models with LFM3

LFM3 represents the next generation of Liquid Foundation Models, consolidating text, vision, audio, omni, and decision models into a shared backbone with approximately 2 billion parameters. Each model can be fine-tuned on a single GPU and deployed locally, addressing specific use cases without requiring large general-purpose models. The architecture leverages hardware-in-the-loop search across 14 device and backend combinations, including phones, Raspberry Pi 5, laptops, and cloud GPUs, optimizing for agentic tasks and long-context inference.

LFM3-1.7B-Base is pre-trained on 22 trillion tokens with a 128K vocabulary and further refined on 3.5 trillion curated tokens, extending the context window to 256K tokens. The model outperforms its predecessor, LFM2.5-2.6B, with 35% fewer parameters, achieving a score of 38.42 on WebArena-Verified compared to 3.02, and surpassing Qwen3.5-9B on long tasks in browsers and business applications. LFM3-VL-2B introduces video input capabilities for visual understanding, reasoning step-by-step, and object placement analysis, raising average performance across 62 vision benchmarks from 57.6 to 71.4 while using fewer parameters than its predecessor.

LFM3-Audio-2B supports voice chat, speech recognition, synthesis, and translation across 18 languages, responding with text, speech, or both. It introduces a new audio encoder and speech codec, improving speech quality (UTMOSv2 3.83 vs. 3.22) and adding audio understanding and tool-calling capabilities. LFM3-Omni-2B integrates text, vision, and audio inputs with text and speech outputs, enabling tasks such as answering spoken questions about images or processing video with soundtracks. It outperforms models up to 10.7 billion parameters on audio-visual benchmarks while maintaining real-time speech streaming via LiquidCodec.

The d1 models focus on structured decision-making, answering multiple questions in a single forward pass with calibrated probabilities. They support tasks like pointing at objects, reading document values, estimating numerical ranges, and selecting applicable options. LFM3 models are designed for specialization, allowing short fine-tuning on custom data to create task-specific experts deployable across devices, private infrastructure, or cloud environments. Example notebooks and robotics adaptations demonstrate fine-tuning for orchestration, navigation, and manipulation tasks, with the manipulation VLA achieving 96.2% success on LIBERO suites and 45.7% on RoboCasa-GR1 tasks.