Liquid AI releases on-device foundation model LFM2 - The Robot Report

Liquid AI has launched LFM2, its latest Liquid Foundation Model designed for on-device deployment, aiming to balance quality, latency, and memory efficiency tailored to specific tasks and hardware. By moving large generative models from cloud servers to local devices such as phones, laptops, cars, and robots, LFM2 offers millisecond latency, offline functionality, and enhanced data privacy. The model features a new hybrid architecture that delivers twice the decode and prefill speed on CPUs compared to Qwen3 and outperforms similarly sized models across benchmarks in knowledge, mathematics, instruction following, and multilingual capabilities. Additionally, LFM2 achieves three times faster training efficiency than its predecessor. LFM2’s architecture includes 16 blocks combining double-gated short-range convolution and grouped query attention, enabling efficient operation on CPUs, GPUs, and NPUs across various devices. Liquid AI provides three model sizes (0.35B, 0.7B, and 1.2B parameters) available under an open