Alpamayo 1 (R1) — W4A8 (ParoQuant)

Flash Vision-Language-Action Inference for Autonomous Driving

W4A8 ParoQuant weights for the language model of z-lab/Alpamayo-R1-10B, used by FlashDrive to accelerate Alpamayo 1 (R1).

ParoQuant (ICLR 2026) is a state-of-the-art INT4 quantizer: learned pairwise rotations suppress activation outliers, closing the accuracy gap with FP16 at near-AWQ speed. Here it quantizes the VLM language model to INT4 weights and INT8 activations (served through vLLM's Marlin kernels); the action expert stays bf16.

Not a standalone model. FlashDrive loads the base checkpoint and fills these quantized weights automatically — you do not load this repository directly.

Usage

import flashdrive

# from_pretrained fetches this -PARO checkpoint automatically
model = flashdrive.from_pretrained("z-lab/Alpamayo-R1-10B")

See the base model card and the FlashDrive repository for the full pipeline.

License

This checkpoint is derived from NVIDIA's Alpamayo weights and is governed by the NVIDIA License, which permits non-commercial use only and extends to derivative works. The FlashDrive inference code is separately released under the MIT License.

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}
}

@article{li2026flashdrive,
  title  = {{FlashDrive: Flash Vision-Language-Action Inference for Autonomous Driving}},
  author = {Li, Zekai and Liang, Yihao and Zhang, Hongfei and Chen, Jian and Liang, Yesheng and Liu, Zhijian},
  year   = {2026}
}