Stage 1 — HSIC inter-tensor bit allocation. Per-tensor sensitivity scored with the Hilbert-Schmidt Independence Criterion against ground-truth actions; greedy per-layer L² allocator assigns a quant type per tensor under a budget.
Stage 2 — Action-Mixed Fisher (AMF) imatrix. Per-element Fisher diagonal under the flow-matching action loss, consumed by llama-quantize for block-level scale optimization.

This variant uses Q3_K_S as the LLM allocator's base type (HSIC sens score), Q3_K vision tower, and the action expert kept at fp16. This checkpoint is the best per-bit Pareto point in the paper in the ActQuant paper.

Bits-per-weight breakdown

Section	BPW
Vision (SigLIP, Q3_K)	3.56
LLM transformer blocks (Q3_K_S base, selective HSIC upgrades)	3.45
Vision + LLM blocks (paper headline)	3.47 ≈ 3.5 BPW
Token embedding (Q8_0)	8.50
Action expert / flow head (fp16)	16.00

The headline BPW excludes the embedding and action-expert tensors, following standard quantization-paper convention: those tensors are not the target of ActQuant's block-wise allocation.

LIBERO closed-loop results

Aggregate success rate across all four LIBERO suites, 500 trials per suite (2 000 total), evaluated through the C++/GGML runtime via the pi05.so pybind11 binding (same code path that runs at deployment):

Suite	Success rate
`libero_spatial`	99.0 %
`libero_object`	99.6 %
`libero_goal`	94.0 %
`libero_10` (long horizon)	92.4 %
Aggregate	96.3 %

Files

File	Purpose
`pi05.gguf`	Merged vision + LLM + action-expert GGUF (the deployable artifact)
`tokenizer.model`	PaliGemma SentencePiece tokenizer
`norm_stats.json`	LIBERO action-quantile normalization stats

Run LIBERO evaluation

bash

huggingface-cli download NU-World-Model-Embodied-AI/ActQuant-Pi05-LIBERO-3.5bpw \
    --local-dir /path/to/eval_dir

# From the ActQuant build tree (build_openpi/) — see paper repo:
cd /path/to/ActQuant
for suite in libero_spatial libero_object libero_goal libero_10; do
    bash tools/pi0.5/run_libero_eval.sh "$suite" 50 5 8 8000 /path/to/eval_dir
done

Reproduce this exact checkpoint

bash

# Stage 0 — Export the LIBERO-finetuned bf16 Pi 0.5 to GGUF
#   (vision Q3_K + everything else preserved)
python tools/pi0.5/export_pi05.py \
    -d /path/to/pi05_libero_finetuned_v044 \
    -o /path/to/pi05_libero_base_gguf \
    --quant_vision q3k

# Stage 0b — Export the standalone bf16 PaliGemma LLM
python tools/pi0.5/export_pi05_llm.py \
    -d /path/to/pi05_libero_finetuned_v044 \
    -o /path/to/pi05_libero_base_gguf/pali_llm_bf16.gguf

# Stage 2 — AMF Fisher imatrix (computed once, reused across BPW points)
python tools/fisher-diag/get_pi05_calib_data.py \
    --output-dir /path/to/calib_data_raw
python tools/fisher-diag/compute_fisher_pi05.py \
    --checkpoint /path/to/pi05_libero_finetuned_v044 \
    --calib-dir  /path/to/calib_data_raw \
    --output     /path/to/pi05_libero_base_gguf/fisher_flow_perweight.gguf \
    --num-gpus 8 --batch-size 6

# Stage 1 + 2 + merge
bash tools/hsic/run_hsic_quant_pi05.sh \
    --base-type Q3_K_S \
    --max-type  Q4_K \
    --score-key sens \
    --num-gpus  8

Output: pi05_q3ks_hsic_sens_v3k.gguf (rename to pi05.gguf for serving).

Other BPW points in the same Pareto curve

This checkpoint is part of the ActQuant Pi 0.5 LIBERO BPW sweep reported in the paper:

BPW (v+L)	Aggregate SR	HF repo
4.0	96.6 %	NU-World-Model-Embodied-AI/ActQuant-Pi05-LIBERO-4bpw
3.5	96.3 %	NU-World-Model-Embodied-AI/ActQuant-Pi05-LIBERO-3.5bpw
3.0	94.8 %	NU-World-Model-Embodied-AI/ActQuant-Pi05-LIBERO-3bpw

All variants are in the ActQuant Collection.

Citation

bibtex

@article{actquant2026,
  title  = {ActQuant: Sub-4-bit Action-Guided Quantization for Vision-Language-Action Models},
  author = {Akbari, Arash and others},
  journal= {arXiv preprint arXiv:2605.24011},
  year   = {2026}
}

License

MIT (inherited from the llama.cpp upstream build infrastructure). Pi 0.5 model weights are bound by the upstream OpenPI license and the underlying PaliGemma terms.