ParoQuant

Pairwise Rotation Quantization for Efficient Reasoning LLM Inference

State-of-the-art INT4 quantization for LLMs. ParoQuant uses learned pairwise rotations to suppress weight outliers, closing the accuracy gap with FP16 while running at near-AWQ speed. Supports NVIDIA GPUs (vLLM, Transformers) and Apple Silicon (MLX).

Quick Start

oMLX (Apple Silicon)

On Apple Silicon, we recommend serving ParoQuant models using oMLX. Refer to the docs for more details.

Installation

# NVIDIA GPU (CUDA 12.9)
pip install "paroquant[vllm]"

# NVIDIA GPU (CUDA 13.0)
pip install "paroquant[vllm]" "vllm==0.19.1" \
  --extra-index-url https://wheels.vllm.ai/0.19.1/cu130 \
  --extra-index-url https://download.pytorch.org/whl/cu130

# Apple Silicon
pip install "paroquant[mlx]"

Pick a model from our Hugging Face collection:

export MODEL=z-lab/Qwen3.5-4B-PARO

Interactive Chat

python -m paroquant.cli.chat --model $MODEL

OpenAI-Compatible API Server

For vLLM, you can directly use vllm serve to serve ParoQuant models:

vllm serve $MODEL --port 8000

For other frameworks:

python -m paroquant.cli.serve --model $MODEL --port 8000

For MLX, add --vlm if you wish to load the VLM components and use the model's multimodal features. For vLLM, VLM components are loaded by default and can be skipped with the server argument --language-model-only.

Docker (NVIDIA GPU)

[!NOTE] The following commands map the local cache directory to the container in order to persist kernel cache across runs. Remove -v ... to disable this behaviour.

# Interactive chat
docker run --pull=always --rm -it --gpus all --ipc=host \
  -v $HOME/.cache/paroquant:/root/.cache/paroquant \
  ghcr.io/z-lab/paroquant:chat --model $MODEL

# API server (port 8000)
docker run --pull=always --rm -it --gpus all --ipc=host -p 8000:8000 \
  -v $HOME/.cache/paroquant:/root/.cache/paroquant \
  ghcr.io/z-lab/paroquant:serve --model $MODEL

Models

All models are available on Hugging Face. Swap the model name in the commands above to try any of them.

Gemma 4

Qwen3.6

Qwen3.5

Qwen3

Llama

Want a model that's not listed? Open an issue and let us know.

Reproduction

[!NOTE] The main branch of this repository is under active development, and reproducibility is not guaranteed. Please use the legacy branch to reproduce results from the paper.

Quantize Your Own Model

git clone https://github.com/z-lab/paroquant && cd paroquant
pip install -e ".[optim,eval]"

# 1. Optimize rotation parameters
experiments/optimize/4bit.sh Qwen/Qwen3-8B

# 2. Export to HF checkpoint (--mode real for INT4, --mode pseudo for FP16)
python -m paroquant.cli.convert \
  --model Qwen/Qwen3-8B \
  --result-dir output/Qwen3-8B \
  --output-path models/Qwen3-8B-PARO

Docker Images

Citation

@inproceedings{liang2026paroquant,
  title     = {{ParoQuant: Pairwise Rotation Quantization for Efficient Reasoning LLM Inference}},
  author    = {Liang, Yesheng and Chen, Haisheng and Zhang, Zihan and Han, Song and Liu, Zhijian},
  booktitle = {International Conference on Learning Representations (ICLR)},
  year      = {2026}
}