Liquid AI today released quantized versions of its LFM2.5 language models optimized for edge deployment. The Q4_0 checkpoints—available in sizes ranging from 230M to 2.6B parameters—use Quantization-Aware Distillation (QAD), a technique that trains quantized student models under the guidance of high-precision teacher models. Across the four released model sizes, QAD recovers between 96.5% and 97.4% of the performance that would otherwise be lost during quantization, substantially outperforming traditional post-training quantization methods. The quantized models maintain the memory footprint and throughput benefits of standard 4-bit quantization while achieving near full-precision accuracy. Benchmarks spanning reasoning, instruction-following, tool use, and agentic capabilities demonstrate that QAD checkpoints match or exceed existing quantization approaches. Speed tests across diverse hardware platforms—MacBook Pro, Samsung Galaxy S26 Ultra, and Raspberry Pi 5—show 3-33% higher decode throughput compared to higher-precision quantized variants, making them practical for real-world edge deployment. The models are immediately available on Hugging Face and compatible with llama.cpp and other runtimes supporting GGUF Q4_0 artifacts. The release addresses a persistent challenge in making capable language models deployable on resource-constrained devices without sacrificing quality—a requirement increasingly critical for on-device AI applications and privacy-focused inference.