Liquid AI ships 4-bit LFM2.5 checkpoints that hold 97% of full-precision accuracy
Liquid AI released Q4_0 checkpoints for four LFM2.5 models — 230M, 350M, 1.2B-Instruct and 2.6B — trained by distilling a high-precision teacher into a quantized student rather than compressing after training, and they retain 96.5% to 97.4% of BF16 accuracy on MMLU-Pro, GPQA Diamond, GSM8K and AIME25. Liquid measured 3-33% higher decode throughput than heavier quantizations on a MacBook Pro, NucBox EVO-X2, Galaxy S26 Ultra and Raspberry Pi 5, so the usual 4-bit quality tax on edge hardware largely disappears. The files run today in llama.cpp or any GGUF Q4_0 runtime.
Source: huggingface.co ↗
The QAD checkpoints retain 97.1%, 96.5%, 97.4%, and 96.6% of their respective BF16 baseline performance.
Liquid AI
Why this matters
- → 4-bit models retain 97% accuracy, eliminating quantization's usual quality penalty.
- → 3-33% faster decode on edge hardware (phone, Pi, MacBook) with same memory footprint.
- → Distill-then-quantize beats compress-after-training; usable today in llama.cpp.
Quantization without compromise