Llama-3.1-8B-Instruct (GGUF) — Unified llama.cpp Quantization Benchmarks

This repository provides a set of GGUF quantizations of Llama-3.1-8B-Instruct, produced with the official llama.cpp toolchain and evaluated under a single, consistent experimental protocol. The goal is to make it practical to choose a GGUF quantization format based on measured trade-offs among quality, model size, and CPU throughput. For the full benchmark evaluation, refer to the paper.

Paper: Which Quantization Should I Use? A Unified Evaluation of llama.cpp Quantization on Llama-3.1-8B-Instruct

What’s in this repo

All files are GGUF versions of the same base checkpoint:

• Reference: FP16 GGUF baseline (used as the evaluation reference)
• Legacy formats: Q4_0, Q4_1, Q5_0, Q5_1, Q8_0
• K-quant formats: Q3_K_S, Q3_K_M, Q3_K_L, Q4_K_S, Q4_K_M, Q5_K_S, Q5_K_M, Q6_K

These are the same families evaluated in the paper’s unified benchmark sweep.

How the quantizations were produced

All quantized models were generated from the same FP16 GGUF input using the official llama.cpp quantizer, varying only the scheme identifier:

./llama-quantize <f16.gguf> <out.gguf> <SCHEME>

Citation

@article{kurt2026quantization,
  title={Which Quantization Should I Use? A Unified Evaluation of llama. cpp Quantization on Llama-3.1-8B-Instruct},
  author={Kurt, Uygar},
  journal={arXiv preprint arXiv:2601.14277},
  year={2026}
}