Liquid LFM 2.5 hits edge with vision and audio

Liquid AI's hybrid architecture bypasses quadratic scaling issues, enabling long-context reasoning with a minimal memory footprint. The 1.2B Instruct model's IFEval performance supports local agentic workflows, while native NPU optimization signals a shift toward specialized hardware for private inference. Multimodal variants like VL-1.6B and Audio-1.5B further enable low-latency vision and voice interfaces on-device without cloud dependencies.