Add ParoQuant (Pairwise Rotation Quantization) Support to MLX-LM

[Revive-Curiosity]

Feature Request

I would like to request support for ParoQuant (Pairwise Rotation Quantization) in MLX-LM.

Background

ParoQuant is a new post-training quantization method introduced at ICLR 2026. It uses pairwise rotation to suppress outliers in weight distributions, making it especially effective for reasoning-heavy LLMs. Compared to standard PTQ, it achieves better accuracy retention while still reducing memory and compute costs.

Why MLX-LM?

MLX-LM already supports uniform, mixed-bit, and affine quantization. Adding ParoQuant would:

• Improve robustness for reasoning-focused models.
• Enable developers to experiment with cutting-edge quantization methods on Apple Silicon.
• Keep MLX aligned with the latest research in efficient LLM deployment.

Suggested Implementation

• Integrate ParoQuant’s pairwise rotation preprocessing step into MLX’s quantization pipeline.
• Provide options for INT4/INT8 precision.
• Allow exporting/importing ParoQuant-quantized weights for Hugging Face compatibility.

References

• ParoQuant paper (ICLR 2026): Liang, Chen, Han, Liu
• GitHub repo: z-lab/paroquant

Thanks for considering this request!

[zhijian-liu]

We’d love to bring native support for ParoQuant into MLX. Looking forward to discussing the best way to move this forward!

[ivanfioravanti]

I'm testing ParoQuant with ton of evals and it's outstanding! @angeloskath it would be amazing having it natively integrated in MLX. Amazing work @zhijian-liu publishing eval results soon! 3 days of evals and still ongoing.

[ivanfioravanti]

I did some evals with Qwen3.5-9B and here results, I have limited max-tokens in some cases, but the important thing is that same limits have been applied on both quantizations:

Benchmark	ParoQuant	4bit
ARC-Challenge	0.956	0.919
GSM8K	0.770	0.605
HumanEval	0.933	0.905
IFEval	0.382	0.172
HellaSwag	0.816	0.792
MMLU	0.794	0.652
GPQA	0.730	0.580

[zhijian-liu]

Thank you, @ivanfioravanti!