Liquid AI unveils lfm3: a suite of 6 multimodal models with enhanced performance

Liquid AI has launched LFM3, a new generation of multimodal foundation models comprising six open-weight, 2B-class models. The family, built on a shared backbone, supports text, vision, audio, omni, and decision tasks, with each model fine-tunable on a single GPU. The backbone integrates 3 operators across 24 layers, including 13 gated short convolutions, 5 sliding-window attention layers, and 6 global attention layers. The model was tuned across 14 device and backend combinations, including a Raspberry Pi 5 and Qualcomm Hexagon NPU. LFM3-1.7B outperforms LFM2.5-2.6B with 35% fewer parameters, achieving scores of 38.42 on WebArena-Verified and 71.39 on BFCL V4. The models are available on Hugging Face, the Liquid AI Playground, and the Liquid API, with support for GGUF, MLX, and ONNX formats. Liquid AI also demonstrated improvements in decision tasks, raising the score on Decision Bench from 83.6 to 85.2.