Liquid AI's QAD Checkpoints Shrink LFM2.5 to Edge Size Without the Usual Quality Tax

2. Liquid AI's QAD Checkpoints Shrink LFM2.5 to Edge Size Without the Usual Quality Tax

On August 19, 2026, Liquid AI published Quantization-Aware Distillation (QAD) Q4_0 GGUF checkpoints for four LFM2.5 models: LFM2.5-230M, LFM2.5-350M, LFM2.5-1.2B-Instruct, and LFM2.5-2.6B. The method distills a high-precision teacher model into a quantized student model, keeping the same memory footprint and throughput as standard Q4_0 while recovering 97.1%, 96.5%, 97.4%, and 96.6% of each model's respective BF16 baseline accuracy. Benchmarks span reasoning, instruction-following, tool use, and agentic tasks including GPQA Diamond, MMLU-Pro, IFEval, and BFCLv4. Hardware targets include MacBook Pro, NucBox EVO-X2, Samsung Galaxy S26 Ultra, and Raspberry Pi 5.

The release sharpens a real competitive pressure point. Post-training quantization has long been the default path to smaller models, but quality loss at 4-bit has been the accepted cost. Liquid AI's QAD checkpoints match Q5_K_M quality at 4-to-33% higher decode throughput for the 230M and 350M models, and match Q4_K_M quality at 3-to-14% higher throughput for the 1.2B and 2.6B. Critically, the QAD Q4_0 checkpoints also match Unsloth's UD-Q4_K_XL on the 230M and 1.2B, a well-regarded external PTQ baseline. For teams choosing between quantization providers, that parity removes one of Unsloth's clearest differentiators on these model sizes.

This fits a broader pattern Liquid AI has been executing since early August 2026: a rapid edge deployment push across the LFM2.5 family, including the LFM2.5-VL-3B vision release on August 12 and the LFM2.5-2.6B local agent deployment post on August 4. The non-transformer architecture is being positioned specifically around inference efficiency on constrained hardware. The next signal to watch: whether QAD techniques extend to the VL-3B vision model, where multimodal edge inference remains an unsolved cost problem for most on-device deployments.

Source: LFM2.5 Q4_0 Checkpoints from Quantization-Aware Distillation