Liquid AI releases distilled 4-bit LFM2.5 checkpoints
Liquid AI has released Q4_0 checkpoints of four LFM2.5 models trained with quantization-aware distillation, keeping the speed of 4-bit GGUFs while retaining roughly 97% of full-precision benchmark scores.
Liquid AI announced updated 4-bit checkpoints for LFM2.5-230M, LFM2.5-350M, LFM2.5-1.2B-Instruct and LFM2.5-2.6B. All four were trained with Quantization-Aware Distillation (QAD), a method that distills a high-precision teacher model into a quantized student. The GGUFs are available now on Hugging Face.
The company compared the new checkpoints against its existing GGUFs produced with post-training quantization (PTQ), using the BF16 GGUF as the ceiling. The test suite covered GPQA Diamond, MMLU-Pro, IFEval, IFBench, Multi-IF and BFCLv4, plus GSM8K for the two smaller models and AIME25 for the two larger ones. Scores are averaged over five runs.
QAD closed 70.6% of the gap between BF16 and Q4_0 for the 230M model, 73.4% for 350M, 65.5% for 1.2B and 48.4% for 2.6B. Measured against their BF16 baselines, the four checkpoints retain 97.1%, 96.5%, 97.4% and 96.6% of performance. The smaller models gained the most because PTQ costs them more quality to begin with.
Q4_0 is the fastest 4-bit format but the weakest on quality. The QAD versions use the same GGUF format and runtime path, so they run at native Q4_0 speed in llama.cpp. Liquid tested throughput on an Apple M5 Max, an AMD Ryzen AI Max+ 395, a Snapdragon 8 Elite Gen 5 in the Galaxy S26 Ultra, and a Raspberry Pi 5.
Liquid says QAD Q4_0 matches Q5_K_M quality within run-to-run variance at 230M and 350M, and comes within half a point at 1.2B, while being smaller and faster. It also reports that the checkpoints beat Q4_K_M and match Unsloth's UD-Q4_K_XL.