Feihua-n1-64M-GGUF

中文版请点击 / Chinese version: README_CN.md

<div align="center">

Feihua-n1-64M-GGUF: A Full-Spectrum Quantization Repository for the Feihua Model

How small is small? The smallest one fits on fifteen 3.5-inch floppy disks.

</div>

Abstract

This repository is the complete llama.cpp GGUF quantization suite for Feihua-n1-64M, a 64M zero-information model of Feihua (废话) literature, spanning seven tiers (Q8_0 / Q4_K_M / IQ4_XS / IQ2_M / IQ2_XXS / IQ1_M / IQ1_S). Every quantization is calibrated with an imatrix computed over 3,000 full-length incoherent articles — our own data calibrating our own quantization, so no good water flows into a stranger's field. The full research report on training, reinforcement, and compression lives in the main repository's model card; quantizations of the pruned variant (49M, 6 layers) are collected separately in Feihua-n1-64M-prune-GGUF. This repository does one thing and one thing only: packing Feihua into smaller boxes.

File Listing

(PPL is measured on the Feihua corpus; lower is better, with the f16 master baseline at 120.5. Bit widths are llama.cpp nominal values. See Sections 3.3/3.4 of the main repository's research report for the full methodology and analysis.)

Quick Start

# 甜点档
llama-server -hf ZZRI/Feihua-n1-64M-GGUF:feihua-v6-IQ4_XS

# 极限档
llama-server -hf ZZRI/Feihua-n1-64M-GGUF:feihua-v6-IQ2_M

# 行为艺术档
llama-server -hf ZZRI/Feihua-n1-64M-GGUF:feihua-v6-IQ1_S

Acknowledgements & License

Conversion and quantization are built on llama.cpp; the base model and training framework are minimind (Apache-2.0); the raw material for the calibration corpus comes from BullshitGenerator. This repository is released under Apache-2.0.

@misc{feihua-n1-64m-gguf,
  title  = {Feihua-n1-64M-GGUF: 废话模型的全谱系量化仓库},
  author = {ZZRI},
  year   = {2026},
  note   = {把废话装进更小的盒子里}
}