File	Purpose	Size
`Mage-Flow-Edit-XPO3-NVFP4.safetensors`	Self-contained Edit transformer	4.65 GB
`qwen3vl_4b_fp8_scaled.safetensors`	Shared scaled-FP8 text encoder	5.24 GB
`Mage-Flow-VAE.safetensors`	Shared Mage VAE	345 MB

The root config.json is included for Hugging Face model/download accounting.

Quick start

Linux, Python 3.11, CUDA 13, and an NVIDIA Blackwell SM120 GPU are required.

bash

hf download ajh-code/Mage-Flow-Edit-XPO3-NVFP4 --local-dir mage-edit-xpo3
cd mage-edit-xpo3
python3.11 -m venv .venv
. .venv/bin/activate
pip install -r requirements.txt
python edit.py reference.jpg "Change the background; keep the subject unchanged." --output edited.png

The promoted profile is 30 steps, CFG 5, image-only fused GELU, calibrated FP4 bridge on all 12 blocks, direct-HND attention on steps 7-29, and exact SDPA fallback on steps 0-6. Every optimization has a CLI disable switch.

Measured performance

GPU · 512-long-edge edit	BF16 hot denoise	XPO3 hot denoise	Speedup
RTX 5080 · Dog background	4.2491 s	2.5709 s	1.65×
RTX 5080 · Fruit tablecloth	4.4839 s	2.6815 s	1.67×
RTX 5060 Ti · Dog background	8.6019 s	4.7614 s	1.81×
RTX 5060 Ti · Fruit tablecloth	9.3282 s	5.1300 s	1.82×

Hot denoise timings are local matched measurements, not universal end-to-end claims. The RTX 5060 Ti runs also used 31-33% less peak allocation than BF16.

Validated scope

Same-seed dog and fruit edits were deterministic across fresh and hot runs.
The standalone package reproduced both approved RTX 5080 composed outputs pixel-exactly.
Each release-only edit recorded 552 direct routes, 168 exact fallback routes, all 12 bridge blocks, and restored every temporary patch.

This is an XPO3 runtime package, not a generic portable quantization format. The bundled native libraries target Linux x86-64, Python 3.11, PyTorch 2.13.0+cu130, and SM120.

MIT applies to the XPO3 package and Mage-derived runtime. The scaled-FP8 Qwen3-VL component and bundled SpargeAttn code are Apache-2.0; see THIRD_PARTY_NOTICES.md.