Liquid AI releases LFM3, 'every modality' small models

Liquid keeps pushing non-transformer-ish architectures for on-device use. LFM3 claims all modalities in compact models. Has anyone benchmarked the previous LFM versions against similarly sized open models for real tasks?