Jetson-PI VLA GGUF Models
GGUF conversions of PI0, PI0.5, and NVIDIA Isaac GR00T N1.7 for deployment with Jetson-PI-Edge, a llama.cpp-based inference engine for real-time vision-language-action models on NVIDIA Jetson and other CUDA-capable platforms.
Model files
| Model | File | Description |
|---|---|---|
| PI0 | pi0/pi_llm.gguf | PI0 language model and action expert, F16 |
| PI0 | pi0/mmproj-model-f16.gguf | Matching PI0 SigLIP vision encoder/projector, F16 |
| PI0.5 | pi05/pi_llm.gguf | PI0.5 language model and action expert, F16 |
| PI0.5 | pi05/mmproj-model-f16.gguf | Matching PI0.5 SigLIP vision encoder/projector, F16 |
| GR00T N1.7 | gr00t-n1d7/gr00t-n1d7-bf16.gguf | Qwen3-VL backbone and GR00T Action Head, BF16 |
| GR00T N1.7 | gr00t-n1d7/mmproj-gr00t-n1d7-bf16.gguf | Matching Cosmos-Reason2/Qwen3-VL vision projector, BF16 |
Always use the main model and mmproj from the same model directory.
Download
huggingface-cli download diantoudefengshan/Jetson-PI-GGUF \
gr00t-n1d7/gr00t-n1d7-bf16.gguf \
gr00t-n1d7/mmproj-gr00t-n1d7-bf16.gguf \
--local-dir Jetson-PI-GGUF
Runtime
Build Jetson-PI-Edge with CUDA support:
git clone https://github.com/PKU-SEC-Lab/Jetson-PI-Edge.git
cd Jetson-PI-Edge
cmake -S . -B build -DGGML_CUDA=ON -DCMAKE_BUILD_TYPE=Release
cmake --build build --target llama-server jetson_pi_gr00t -j
Start GR00T server mode (do not set PI_MODEL):
./build/bin/llama-server \
-m /path/to/gr00t-n1d7/gr00t-n1d7-bf16.gguf \
--mmproj /path/to/gr00t-n1d7/mmproj-gr00t-n1d7-bf16.gguf \
-ngl 99 --host 0.0.0.0 --port 8080
PI0 and PI0.5 server commands and the full foreground HTTP API are documented in the Jetson-PI-Edge README.
GR00T conversion and validation
The GR00T files were converted from nvidia/GR00T-N1.7-3B and
nvidia/Cosmos-Reason2-2B with the GR00T converter in Jetson-PI-Edge. The main
GGUF contains 715 tensors and the mmproj contains 316 tensors. They passed the
repository's metadata, tensor-name/shape, 16-backbone-block, vision-split, and
end-to-end action checks. With aligned inputs and initial denoising noise, the
Jetson-PI normalized [40, 132] action achieved cosine similarity 0.999968648
against the official NVIDIA PyTorch implementation.
SHA-256:
8cd6af6b33d90d9b6b6fdbf419dbdbc263393dbcb9ad71bbd226889b4285d573 gr00t-n1d7-bf16.gguf
f9c442143857ee01bf95192516de8169af7e0e136d410942a3eeee316c06bdd4 mmproj-gr00t-n1d7-bf16.gguf
Intended use and license
These models are intended for research on local and onboard VLA inference, robot-control systems, deployment optimization, and reproducibility. They are not general-purpose chat models. The converted weights remain subject to their respective upstream model licenses and terms; review those terms before redistribution or deployment. Jetson-PI-Edge source code is distributed under its repository license.
Acknowledgments
This work builds on OpenPI, NVIDIA Isaac-GR00T, Cosmos-Reason2, and llama.cpp.
Citation
@article{yang2026jetson,
title={Jetson-PI: Towards Onboard Real-Time Robot Control via Foresight-Aligned Asynchronous Inference},
author={Yang, Zebin and Wang, Qi and Wang, Yunhe and Guo, Xiurui and Yu, Bo and Liu, Shaoshan and Xu, Jiafeng and Dong, Hao and Li, Meng},
journal={arXiv preprint arXiv:2607.12659},
year={2026}
}