Feihua-n1-64M-prune-GGUF

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

<div align="center">

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

A quarter of the brain removed, nearly half the size shed; the smallest one can be carried away on twelve 3.5-inch floppy disks.

</div>

Abstract

This repository hosts the complete set of llama.cpp GGUF quantizations of Feihua-n1-64M-prune — a 49.16M-parameter, 6-layer Feihua (废话) literature model — in seven tiers (Q8_0 / Q4_K_M / IQ4_XS / IQ2_M / IQ2_XXS / IQ1_M / IQ1_S). All quantizations are imatrix-calibrated against a corpus of 3,000 long-form articles of unadulterated nonsense. The full research report on pruning and recovery fine-tuning lives on the main model card; quantizations of the unpruned sibling (64M, 8 layers) can be found in Feihua-n1-64M-GGUF. This repository's mission remains singular: to pack a brain-trimmed nonsense engine into ever smaller boxes.

File List

(PPL measured on the Feihua corpus — lower is better; the f16 master baseline is 120.5. IQ1_S and IQ1_M both come in at 17MB: at this size, what remains in the file is mostly the bedrock of embedding tables and metadata, and no further drop in bit width can squeeze out any more bytes.)

Quick Start

# sweet spot
llama-server -hf ZZRI/Feihua-n1-64M-prune-GGUF:feihua-v6-prune6-heal-IQ4_XS

# at the limit (still Feihua at 21MB)
llama-server -hf ZZRI/Feihua-n1-64M-prune-GGUF:feihua-v6-prune6-heal-IQ2_M

# performance art (17MB, poetry only)
llama-server -hf ZZRI/Feihua-n1-64M-prune-GGUF:feihua-v6-prune6-heal-IQ1_S

Acknowledgements & License

Conversion and quantization are built on llama.cpp; the pruning method is ShortGPT (arXiv:2403.03853); the base model and training framework are minimind (Apache-2.0); the calibration corpus is derived from BullshitGenerator. This repository is released under Apache-2.0.

@misc{feihua-n1-64m-prune-gguf,
  title  = {Feihua-n1-64M-prune-GGUF: A Full-Spectrum Quantization Repository for the Pruned Feihua Model},
  author = {ZZRI},
  year   = {2026},
  note   = {A quarter of the brain gone, nearly half the size shed, and not a single gram of Feihua lost}
}