Liquid AI just made a strong case that aggressive quantization doesn’t have to gut model quality. Their new LFM2.5 Q4_0 checkpoints, trained with Quantization-Aware Distillation (QAD), recover 97% of the average accuracy that standard post-training quantization (PTQ) typically destroys when compressing models to 4-bit. The technique distills a high-precision BF16 teacher into a quantized student, and the results hold across all four LFM2.5 sizes: 230M, 350M, 1.2B-Instruct, and 2.6B.
Across a benchmark suite spanning GPQA Diamond, MMLU-Pro, IFEval, IFBench, Multi-IF, and BFCLv4 — plus GSM8K for the smaller models and AIME25 for the larger pair — the QAD checkpoints retained between 96.5% and 97.4% of their BF16 baseline performance. Those numbers matter because PTQ Q4_0 typically leaves visible scars on reasoning and instruction-following tasks. QAD appears to close most of that gap while keeping the memory footprint and decode speed of native Q4_0 GGUFs untouched.
The on-device numbers are equally interesting. On a Raspberry Pi 5 running Arm CPU inference, the 230M and 350M QAD checkpoints match Q5_K_M quality within evaluation variance while delivering 4-33% higher decode throughput. The 1.2B and 2.6B models match Q4_K_M quality at 3-14% faster speeds. They also hold their own against Unsloth’s UD-Q4_K_XL, a strong external PTQ checkpoint, on the models where that comparison applies.
This is a meaningful shift for edge deployment. For years, developers have accepted a brutal tradeoff: shrink the model, lose the reasoning. Liquid AI’s QAD approach suggests that tradeoff was partly an artifact of how quantization was being done, not an inherent limit of 4-bit precision. The checkpoints are live on Hugging Face now, and they drop into any llama.cpp-compatible runtime. If QAD becomes standard practice across other model families, the floor for usable on-device AI just got a lot higher.
