← Back to the wire

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

AchievementModelAug 19, 2026

Liquid AI released LFM2.5 Q4_0 checkpoints trained with Quantization-Aware Distillation, recovering 97% of BF16 average accuracy lost to quantization. The four models—LFM2.5-230M, LFM2.5-350M, LFM2.5-1.2B-Instruct, and LFM2.5-2.6B—retain the low memory footprint and high throughput of native Q4_0 GGUFs. Benchmarked across reasoning, instruction-following, tool use, and agentic tasks, the QAD checkpoints match or exceed comparable post-training quantization quality at higher decode throughput on edge hardware.

Receipt № 14651 source · awaiting confirmation ◐

Evidence

1source· awaiting independent confirmation

No score is assigned. Sources and their independence are shown in the citation chain below.

Citation chain · 1 source

01medPRIMARY
LFM2.5-1.2B-InstructModelLFM2.5-230MModelLFM2.5-350MModelLFM2.5-2.6BModelLFM2.5 Q4_0Model
Canonical: https://huggingface.co/blog/LiquidAI/qad