LFM2.5 Q4\_0 Checkpoints from Quantization-Aware Distillation

Liquid AI has put out new Q4_0 GGUF checkpoints for four LFM2.5 models: 230M, 350M, 1.2B-Instruct, and 2.6B. The pitch is simple and pretty appealing: you get the usual Q4_0 memory footprint and throughput, but without taking the normal quality haircut that comes with quantization.

The trick is Quantization-Aware Distillation, or QAD. Instead of quantizing first and hoping for the best, the company distills a high-precision teacher into a quantized student. In Liquid AI’s numbers, that recovers 97% of the BF16 average accuracy lost to quantization. That is the whole point of the release, and it’s a strong one.

On the benchmark suite, the company compares these QAD checkpoints with its earlier post-training quantized GGUFs across reasoning, instruction-following, tool use, and agentic tasks: GPQA Diamond, MMLU-Pro, IFEval, IFBench, Multi-IF, and BFCLv4. It also adds GSM8K for the smaller two models and AIME25 for the larger two, with results averaged across five repeats. The QAD checkpoints keep 97.1%, 96.5%, 97.4%, and 96.6% of their BF16 baseline performance, depending on the model.

The edge-hardware numbers are just as relevant. Liquid AI profiled decode throughput on a MacBook Pro, a NucBox EVO-X2, a Samsung Galaxy S26 Ultra, and a Raspberry Pi 5. The smaller two models, 230M and 350M, matched Q5_K_M quality within evaluation variance while running 4% to 33% faster. The 1.2B and 2.6B versions matched Q4_K_M quality at 3% to 14% higher throughput. Where applicable, they also match Unsloth’s UD-Q4_K_XL, which is a useful outside check.

These files are already on Hugging Face, and they’re meant to run with llama.cpp or any GGUF Q4_0 runtime. That makes the release feel less like a benchmark trophy and more like a practical nudge: if smaller quantized models have been just a bit too rough for real use, Liquid AI is trying to remove that excuse.