README.md

---
license: other
license_name: nvidia-license
license_link: https://huggingface.co/nvidia/Alpamayo-1.5-10B/blob/main/LICENSE
pipeline_tag: robotics
language:
- en
inference: false
base_model:
- nvidia/Alpamayo-R1-10B
new_version: z-lab/Alpamayo-1.5-10B
datasets:
- nvidia/PhysicalAI-Autonomous-Vehicles
- nvidia/PhysicalAI-Autonomous-Vehicles-NuRec
tags:
- flashdrive
- autonomous-driving
- vision-language-action
- alpamayo
---

# Alpamayo 1 (R1) (FlashDrive)

**Flash Vision-Language-Action Inference for Autonomous Driving**

![Paper](https://img.shields.io/badge/arXiv-coming%20soon-b31b1b.svg)
[![GitHub](https://img.shields.io/badge/GitHub-FlashDrive-181717?logo=github)](https://github.com/z-lab/flashdrive)
[![Blog](https://img.shields.io/badge/Blog-FlashDrive-blue)](https://z-lab.ai/projects/flashdrive/)
[![Models](https://img.shields.io/badge/%F0%9F%A4%97-Models-yellow)](https://huggingface.co/collections/z-lab/flashdrive)

[FlashDrive](https://github.com/z-lab/flashdrive) accelerates [Alpamayo 1 (R1)](https://huggingface.co/nvidia/Alpamayo-R1-10B) — one of NVIDIA's 10B-parameter vision-language-action models for autonomous driving — by **4.5× with no loss in accuracy**, through streaming inference, [DFlash](https://github.com/z-lab/dflash) speculative reasoning, [ParoQuant](https://github.com/z-lab/paroquant) W4A8 quantization, adaptive action caching, and `torch.compile`.

This repository mirrors the weights of [nvidia/Alpamayo-R1-10B](https://huggingface.co/nvidia/Alpamayo-R1-10B) and is the **base checkpoint** of the FlashDrive stack. Loading it pulls the derived companions automatically:

| Checkpoint | Contents |
|---|---|
| [z-lab/Alpamayo-R1-10B-PARO](https://huggingface.co/z-lab/Alpamayo-R1-10B-PARO) | W4A8 ([ParoQuant](https://github.com/z-lab/paroquant)) language-model weights |
| [z-lab/Alpamayo-R1-10B-DFlash](https://huggingface.co/z-lab/Alpamayo-R1-10B-DFlash) | [DFlash](https://github.com/z-lab/dflash) block-diffusion draft model |

## Usage

Install [FlashDrive](https://github.com/z-lab/flashdrive), then load this base checkpoint — the `-PARO` and `-DFlash` companions are fetched automatically:

```python
import flashdrive

model = flashdrive.from_pretrained("z-lab/Alpamayo-R1-10B")

pred_xyz, pred_rot = model.sample_trajectories_streaming(data)
```

The first call per stream only prefills the KV cache and returns `(None, None)`; every later window returns trajectories. For an end-to-end benchmark on a PhysicalAI-AV clip:

```bash
python scripts/infer.py --model-path z-lab/Alpamayo-R1-10B
```

## Performance

On a single RTX PRO 6000, averaged over 100 PhysicalAI-AV clips, FlashDrive runs Alpamayo 1 (R1) at **4.5× lower latency** (704 → 155 ms per window) while minADE improves from 1.869 to 1.662. See the [repository](https://github.com/z-lab/flashdrive#performance) for the full benchmark.

## License

The Alpamayo weights in this repository are released by NVIDIA under the [NVIDIA License](https://huggingface.co/nvidia/Alpamayo-1.5-10B/blob/main/LICENSE), which permits **non-commercial use only** and extends to derivative works. The [FlashDrive](https://github.com/z-lab/flashdrive) inference code is separately released under the [MIT License](https://github.com/z-lab/flashdrive/blob/main/LICENSE).

## Citation

```bibtex
@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}
}
```