GGUF conversion of 0xAnkitSingh/GR00T-N1.6-LIBERO (a LIBERO finetune of nvidia/GR00T-N1.6-3B) for inference with vla.cpp, a lightweight C++ inference engine for Vision-Language-Action models built on top of llama.cpp.

GR00T-N1.6 pairs an Eagle3-VL backbone (SigLIP2 vision tower + a Qwen2-2B LLM) with an AlternateVLDiT flow-matching action head, replayed in 16-step action chunks with q01/q99 normalisation. The vision tower is baked into the combined GGUF, so no separate mmproj file is needed.

Unlike GR00T-N1.5 / N1.7 — whose tokenizers live in public HF repos (lerobot/eagle2hg-processor-groot-n1p5, nvidia/Cosmos-Reason2-2B) — N1.6's Eagle3-VL tokenizer has no standalone public repo, so it is bundled in this repo (see Tokenizer).

Files

File	Size	Description
`gr00tn1d6-libero.gguf`	8.5 GiB	Combined VLA model — Eagle3-VL backbone (SigLIP2 + Qwen2-2B) + AlternateVLDiT action head + arch config, BF16
`dataset_statistics.json`	—	Action/state normalisation stats (required by the client)
`tokenizer_config.json`, `vocab.json`, `merges.txt`, `special_tokens_map.json`, `added_tokens.json`, `chat_template.json`	—	Eagle3-VL (`Qwen2TokenizerFast`) tokenizer — pass `--tokenizer` at this directory

Usage

bash

# Terminal 1 — serve (use the CUDA build for inference). No mmproj argument.
VLA_GR00T_BF16_WEIGHTS=1 VLA_GR00T_EMBODIMENT=libero_panda \
    ./build-cuda/vla-server --bind tcp://*:5566 \
    gr00tn1d6-libero.gguf

# Terminal 2 — drive a LIBERO episode (inside the LIBERO uv venv)
python eval/client/run_sim_client_direct.py \
    --arch gr00t_n1_6 \
    --task libero_object --task-id 0 --n-episodes 10 \
    --stats-json dataset_statistics.json \
    --tokenizer . \
    --vla-addr tcp://localhost:5566

Notes:

Set VLA_GR00T_EMBODIMENT=libero_panda and VLA_GR00T_BF16_WEIGHTS=1.
Pass --stats-json dataset_statistics.json (action/state un-normalisation).
Pass --tokenizer . (or the path to this directory) — N1.6's tokenizer is bundled here, not fetched from the Hub.
The client uses --n-action-steps 16 for this checkpoint.

Tokenizer

N1.6's vision-language backbone is Eagle3-VL, whose text tokenizer is a Qwen2TokenizerFast (vocab 151643) plus GR00T-specific added tokens. Neither the base model (nvidia/GR00T-N1.6-3B) nor the LIBERO finetune (0xAnkitSingh/GR00T-N1.6-LIBERO) ships the tokenizer files, and the upstream Eagle3-VL processor repo it references is gated. The six tokenizer files are therefore vendored in this repo; point the client at them with --tokenizer.

Benchmark

Full libero_object sweep (10 tasks × 20 episodes = 200 episodes):

Hardware	n_act	Success rate	client/step	client/call	Peak mem
RTX 3060 (sm_86)	16	86.5%	10.29 ms	165 ms	6048 MiB VRAM
Jetson AGX Orin (sm_87)	16	90.0%	26.70 ms	427 ms	1341 MiB RAM

GR00T-N1.6's ~6 GiB all-resident weight footprint OOMs the Jetson Orin Nano 8 GB unified pool, so it is not in the Orin Nano sweep.

License

Weights follow the upstream license of 0xAnkitSingh/GR00T-N1.6-LIBERO and its base nvidia/GR00T-N1.6-3B (NVIDIA license — review and accept it before use). The vla.cpp conversion tooling and inference engine are Apache-2.0-licensed.

vrfai/gr00tn1d6-libero-gguf

GR00T-N1.6 — LIBERO (GGUF for vla.cpp)

Files

Usage

Tokenizer

Benchmark

License

vrfai/
gr00tn1d6-libero-gguf