The next generation of on-device AI
LFM2 by Liquid AI is a new class of open foundation models designed for on-device speed and efficiency. Its hybrid architecture delivers 2x faster CPU performance than Qwen3 and SOTA results in a tiny footprint.
LFM2 delivers 2x faster decode and prefill performance than Qwen3 on CPU with hybrid architecture optimized for on-device execution. Models run efficiently on smartphones, laptops, vehicles, and embedded systems with millisecond latency and data sovereignty.
LFM2 evaluated across seven languages (Arabic, French, German, Spanish, Japanese, Korean, Chinese) on multilingual benchmarks (MMMLU, MGSM). Outperforms Qwen3 and Gemma 3 on knowledge, mathematics, and multilingual reasoning tasks at comparable parameter counts.
LFM2 significantly outperforms similarly-sized models on instruction-following (IFEval, IFBench) and function-calling benchmarks. Available in 350M, 700M, and 1.2B parameter sizes, enabling efficient agent deployment across diverse hardware with 3x faster training than prior generation.
LFM2 weights are freely available on Hugging Face, llama.cpp, and OpenRouter. Free commercial use for organizations under $10M annual revenue; enterprises can negotiate custom licensing terms directly with sales team.
LEAP platform enables end-to-end customization and deployment workflow. LFM2 architecture was discovered via STAR neural architecture search optimizing for quality, memory, and latency on target hardware like Qualcomm Snapdragon embedded SoCs.