今天,我们发布 QAD Q4_0 GGUF 文件。这些是针对 LFM2.5-230M、LFM2.5-350M、LFM2.5-1.2B-Instruct 和 LFM2.5-2.6B 的更新版 4-bit 检查点。它们让开发者能够以 Q4_0 的内存占用和速度运行 LFM2.5 模型,而不会出现通常的质量下降:
- 采用量化感知蒸馏(QAD)训练:将高精度教师模型蒸馏到量化学生模型中
- 与原生 Q4_0 相同的内存和速度:它们保持了 Q4_0 GGUF 的低内存占用和高吞吐量
- 恢复率:量化导致的 BF16 平均精度损失中,有 97% 得以恢复
基准测试结果
对于全部四个模型,我们将它们发布的、通过训练后量化(PTQ)生成的 GGUF 文件与训练得到的 QAD Q4_0 检查点,在一套涵盖推理、指令遵循、工具使用和智能体能力的基准测试套件上进行了对比:GPQA Diamond、MMLU-Pro、IFEval、IFBench、Multi-IF 和 BFCLv4。BF16 GGUF 作为该格式内的性能上限参考。我们还额外加入了一项规模匹配的数学评测:LFM2.5-230M 和 LFM2.5-350M 使用 GSM8K,LFM2.5-1.2B-Instruct 和 LFM2.5-2.6B 使用 AIME25。我们报告五次重复实验的平均值。
在全部四个模型上,QAD 都显著提升了 Q4_0 检查点的性能。QAD 检查点分别保留了各自 BF16 基线性能的 97.1%、96.5%、97.4% 和 96.6%。
真实边缘硬件上的速度与体积
我们在四个目标设备上测量了 LFM2.5-230M、LFM2.5-350M、LFM2.5-1.2B-Instruct 和 LFM2.5-2.6B 这四个模型的解码吞吐量:MacBook Pro、NucBox EVO-X2、三星 Galaxy S26 Ultra 和 Raspberry Pi 5。MacBook Pro 和 NucBox 使用 GPU 推理,而三星和 Raspberry Pi 使用 Arm CPU 推理。在完成性能分析的情况下,BF16 和 F16 作为全精度参考一并展示。
230M 和 350M 的 QAD Q4_0 检查点在评估误差范围内达到了 Q5_K_M 的质量水平,同时解码吞吐量高出 4-33%。1.2B 和 2.6B 的 QAD Q4_0 检查点达到了 Q4_K_M 的质量水平,同时吞吐量高出 3-14%。QAD Q4_0 检查点还匹配了 Unsloth 的 UD-Q4_K_XL(在适用情况下,即针对 230M 和 1.2B 模型),这是一个强大的外部训练后量化检查点。
如何使用 QAD GGUF 文件
使用 llama.cpp 或任何支持 GGUF Q4_0 工件的运行时来使用这些文件。
llama-cli -hf LiquidAI/LFM2.5-350M \
--hf-file LFM2.5-350M-QAD-Q4_0.gguf \
-p "What is C. elegans?"
开始使用 QAD GGUF 文件
QAD GGUF 格式模型今天已在 Hugging Face 上线:LFM2.5-230M、LFM2.5-350M、LFM2.5-1.2B-Instruct 和 LFM2.5-2.6B。
我们非常期待看到你的创作。
引用
如需引用,请使用以下参考文献或 BibTeX:
Liquid AI, "LFM2.5 Q4_0: Quantization-Aware Distillation for Edge Deployment", Liquid AI Blog, Aug 2026.
或使用 BibTeX 引用
@article{liquidAI2026Q40,
author = {Liquid AI},
title = {LFM2.5 Q4_0: Quantization-Aware Distillation for Edge Deployment},
journal = {Liquid AI Blog},
year = {2026},
note = {www.liquid.ai/blog/qad},
}
Today, we release QAD Q4_0 GGUFs. These are updated 4-bit checkpoints for LFM2.5-230M, LFM2.5-350M, LFM2.5-1.2B-Instruct, and LFM2.5-2.6B. They allow developers to run LFM2.5 models at Q4_0 memory and speed without the usual quality drop:
- Trained with Quantization-Aware Distillation (QAD): a high-precision teacher model is distilled into a quantized student model
- Same memory and speed as native Q4_0: They keep the low memory footprint and high throughput of Q4_0 GGUFs
- Recovery: 97% of their BF16 average accuracy lost to quantization is recovered
Benchmark results
For all four models, we compare their released GGUFs produced with post-training quantization (PTQ) against the trained QAD Q4_0 checkpoints on a benchmark suite spanning reasoning, instruction-following, tool use, and agentic capabilities: GPQA Diamond, MMLU-Pro, IFEval, IFBench, Multi-IF, and BFCLv4. The BF16 GGUF serves as the in-format ceiling. We also add one scale-appropriate math evaluation: GSM8K for LFM2.5-230M and LFM2.5-350M, and AIME25 for LFM2.5-1.2B-Instruct and LFM2.5-2.6B. We report the mean across five repeats.
Across all four models, QAD substantially improves the Q4_0 checkpoint. The QAD checkpoints retain 97.1%, 96.5%, 97.4%, and 96.6% of their respective BF16 baseline performance.
Speed and size on real edge hardware
We measure decode throughput for the four models LFM2.5-230M, LFM2.5-350M, LFM2.5-1.2B-Instruct, and LFM2.5-2.6B across four targets: MacBook Pro, NucBox EVO-X2, Samsung Galaxy S26 Ultra, and Raspberry Pi 5. MacBook Pro and NucBox use GPU inference, while Samsung and Raspberry Pi use Arm CPU inference. BF16 and F16 are shown as full-precision references where profiled.
The 230M and 350M QAD Q4_0 checkpoints match Q5_K_M quality within evaluation variance at a 4-33% higher decode throughput. The 1.2B and 2.6B QAD Q4_0 checkpoints match Q4_K_M quality at a 3-14% higher throughput. The QAD Q4_0 checkpoints also match Unsloth's UD-Q4_K_XL (where applicable, for the 230M and 1.2B), a strong external post-training quantization checkpoint.
How to use QAD GGUFs
Use the files with llama.cpp or any runtime that supports GGUF Q4_0 artifacts.
llama-cli -hf LiquidAI/LFM2.5-350M \
--hf-file LFM2.5-350M-QAD-Q4_0.gguf \
-p "What is C. elegans?"
Get Started with QAD GGUFs
The QAD GGUFs are available on Hugging Face today: LFM2.5-230M, LFM2.5-350M, LFM2.5-1.2B-Instruct, and LFM2.5-2.6B.
We can't wait to see what you build.
Citation
For citations, please use the following reference or BibTeX:
Liquid AI, "LFM2.5 Q4_0: Quantization-Aware Distillation for Edge Deployment", Liquid AI Blog, Aug 2026.
Or use the BibTeX citation
@article{liquidAI2026Q40,
author = {Liquid AI},
title = {LFM2.5 Q4_0: Quantization-Aware Distillation for Edge Deployment},
journal = {Liquid AI Blog},
year = {2026},
note = {www.liquid.ai/blog/qad},
}




