Liquid AI has released Q4_0 GGUF checkpoints for its LFM2.5 model family, utilizing a new training method called Quantization-Aware Distillation (QAD). This approach distills high-precision teacher models into quantized student models to recover accuracy lost during standard post-training quantization.

  • The QAD checkpoints retain between 96.5% and 97.4% of the BF16 baseline performance across four model sizes: LFM2.5-230M, LFM2.5-350M, LFM2.5-1.2B-Instruct, and LFM2.5-2.6B.
  • On edge hardware including MacBook Pro, NucBox EVO-X2, Samsung Galaxy S26 Ultra, and Raspberry Pi 5, the checkpoints match higher-quality quantization levels (Q5_K_M or Q4_K_M) while offering 3-33% higher decode throughput.
  • The models are available on Hugging Face and compatible with llama.cpp and other runtimes supporting GGUF Q4_0 artifacts.

This release enables users to deploy LFM2.5 models on resource-constrained edge devices with performance comparable to larger quantization formats without sacrificing speed.