Instructions to use JiangFeng12/ReaZIT-NPH with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Notebooks
Google Colab
Kaggle
Local Apps Settings

How to use JiangFeng12/ReaZIT-NPH with llama.cpp:

Install (macOS, Linux)

curl -LsSf https://llama.app/install.sh | sh
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf JiangFeng12/ReaZIT-NPH:Q8_0
# Run inference directly in the terminal:
llama cli -hf JiangFeng12/ReaZIT-NPH:Q8_0

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf JiangFeng12/ReaZIT-NPH:Q8_0
# Run inference directly in the terminal:
llama cli -hf JiangFeng12/ReaZIT-NPH:Q8_0

Build from source code

git clone https://github.com/ggml-org/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-app
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama serve -hf JiangFeng12/ReaZIT-NPH:Q8_0
# Run inference directly in the terminal:
./build/bin/llama cli -hf JiangFeng12/ReaZIT-NPH:Q8_0

Use Docker

docker model run hf.co/JiangFeng12/ReaZIT-NPH:Q8_0

Llama new
LM Studio
Jan
Ollama
How to use JiangFeng12/ReaZIT-NPH with Ollama:
```
ollama run hf.co/JiangFeng12/ReaZIT-NPH:Q8_0
```
Unsloth Desktop

How to use JiangFeng12/ReaZIT-NPH with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf JiangFeng12/ReaZIT-NPH:Q8_0

Configure the model in Pi

# Install Pi:
npm install -g @earendil-works/pi-coding-agent
# Add to ~/.pi/agent/models.json:
{
  "providers": {
    "llama-cpp": {
      "baseUrl": "http://localhost:8080/v1",
      "api": "openai-completions",
      "apiKey": "none",
      "models": [
        {
          "id": "JiangFeng12/ReaZIT-NPH:Q8_0"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use JiangFeng12/ReaZIT-NPH with Docker Model Runner:
```
docker model run hf.co/JiangFeng12/ReaZIT-NPH:Q8_0
```

Lemonade

How to use JiangFeng12/ReaZIT-NPH with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull JiangFeng12/ReaZIT-NPH:Q8_0

Run and chat with the model

lemonade run user.ReaZIT-NPH-Q8_0

List all available models

lemonade list

Hermes Agent

How to use JiangFeng12/ReaZIT-NPH with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf JiangFeng12/ReaZIT-NPH:Q8_0

Configure Hermes

# Install Hermes:
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
hermes setup
# Point Hermes at the local server:
hermes config set model.provider custom
hermes config set model.base_url http://127.0.0.1:8080/v1
hermes config set model.default JiangFeng12/ReaZIT-NPH:Q8_0

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use JiangFeng12/ReaZIT-NPH with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf JiangFeng12/ReaZIT-NPH:Q8_0

Configure OpenClaw

# Install OpenClaw:
npm install -g openclaw@latest
# Register the local server and set it as the default model:
openclaw onboard --non-interactive --mode local \
  --auth-choice custom-api-key \
  --custom-base-url http://127.0.0.1:8080/v1 \
  --custom-model-id "JiangFeng12/ReaZIT-NPH:Q8_0" \
  --custom-provider-id llama-cpp \
  --custom-compatibility openai \
  --custom-text-input \
  --accept-risk \
  --skip-health

Run OpenClaw

openclaw agent --local --agent main --message "Hello from Hugging Face"

ReaZIT-NPH · ZIT-HEXAGON 0.30.1

Español (México)

Pesos convertidos para NPH / ZIT-HEXAGON, no el checkpoint fuente original. ReaZIT v1.0 fue creado por SassyAI: modelo original en Civitai. Gracias a SassyAI por permitir esta publicación de la conversión, con crédito y enlace a su perfil. La autorización de publicación fue comunicada por JiangFeng12; este repositorio no es una publicación oficial de SassyAI.

Solo uso no comercial de esta distribución convertida. No vender los pesos ni ponerlos detrás de un muro de pago. Se deben conservar el crédito a SassyAI y los enlaces al perfil y modelo originales. Aplican también los términos originales de ReaZIT y de su modelo base Z-Image Turbo (Apache-2.0). Los pesos no se publican bajo MIT. El código de conversión y la app están en Xiemox2/zit-hexagon-nph bajo MIT (con avisos de terceros). Consulta LICENSE.md y NOTICE.md para el alcance por archivo.

Archivos y uso

reazit_v10_fp8_e4m3fn_scaled.safetensors: DiT ReaZIT v1.0 convertido a FP8 E4M3 con escala por tensor, formato compatible con NPH. No es el checkpoint FP16 original.
Qwen3-4B-Q8_0.gguf: codificador de texto Qwen3-4B cuantizado a Q8_0, dependencia necesaria del flujo ReaZIT; no es un modelo SDXL.
ae.safetensors: VAE estándar de Z-Image Turbo (idéntico por SHA-256 al de Comfy-Org), dependencia del flujo.

Instala la APK recomendada ZIT-HEXAGON v0.30.1-stable.
Descarga los tres archivos y verifica sha256sum -c SHA256SUMS.
Puedes ponerlos en Download/ZITHEXAGON/sdcpp/. En Modelos (NPH) importa/selecciona cada archivo en su rol: DiT, TE y VAE. No hace falta cambiar el nombre del DiT si lo seleccionas explícitamente.
Selecciona ReaZIT FP8 (NPH). Punto de partida: Euler, 8 pasos, CFG 1, 1024×1024; ajusta según la RAM y el dispositivo. Referencia probada del proyecto: SM8850 / HTP v81.
En 0.30.1, si falla una segunda imagen con LoRA y DiT residente, usa Liberar / sal de Extremo antes de generar otra vez. Las builds 0.30.2–0.30.5 no se recomiendan para LoRAs en el dispositivo de referencia.

Estos son los pesos del camino NPH (sd.cpp/ggml-hexagon), no los antiguos contextos QNN dit_*.bin ni el ZIP v81. No se incluyen modelos SDXL de pago, LoRAs, checkpoints fuente ni APKs.

English

Converted weights for NPH / ZIT-HEXAGON, not the original source checkpoint. Original ReaZIT v1.0 author: SassyAI — original Civitai model. Thank you to SassyAI for permitting publication of this conversion with author credit and a profile link. Publication permission was reported by JiangFeng12; this is not an official SassyAI release.

Non-commercial use only for this converted distribution. Do not sell the weights or lock them behind a paywall. Keep SassyAI's credit and links to her profile and original model. The original ReaZIT terms and the Z-Image Turbo base-model terms (Apache-2.0) also apply. The weights are not MIT-licensed. Conversion code and app: Xiemox2/zit-hexagon-nph, MIT, subject to third-party notices. See LICENSE.md and NOTICE.md for per-component terms.

How to use

Install the recommended ZIT-HEXAGON v0.30.1-stable APK.
Download all three files and check sha256sum -c SHA256SUMS.
Optionally place them in Download/ZITHEXAGON/sdcpp/. In Models (NPH) explicitly import/select the ReaZIT FP8 file as DiT, Qwen3-4B-Q8_0.gguf as TE, and ae.safetensors as VAE. Explicit selection does not require renaming the DiT.
Select ReaZIT FP8 (NPH). Starting settings: Euler, 8 steps, CFG 1, 1024×1024, subject to device/RAM support. Project reference device: SM8850 / HTP v81.
For the known 0.30.1 second-image LoRA issue with resident DiT, Release the model / leave Extreme mode before generating again. Builds 0.30.2–0.30.5 are not recommended for LoRA generation on the reference device.

The DiT is FP8 E4M3 with per-tensor scales. The Q8_0 text encoder and standard VAE are runtime dependencies, not SDXL models. This release does not include the original source checkpoint, old QNN contexts/ZIPs, paid SDXL models, LoRAs or APKs.

Tamaños / Sizes and SHA-256

Hashes are the SHA-256 object identifiers verified against the source and destination Hugging Face LFS metadata; no local model download was needed for the server-side copy.

File	Bytes	SHA-256
`reazit_v10_fp8_e4m3fn_scaled.safetensors`	6,158,115,074	`cdd291b39df35238fb80dd245a3bc338affdbe48b76621272ce9eb2baab2dec0`
`Qwen3-4B-Q8_0.gguf`	4,280,404,928	`0a2cee0a56ce2c8499f9ac607a10b9ef06a4e71fbe0a781ffa5b2f54c5d9ab8e`
`ae.safetensors`	335,304,388	`afc8e28272cd15db3919bacdb6918ce9c1ed22e96cb12c4d5ed0fba823529e38`

Total runtime weights: 10,773,824,390 bytes (~10.03 GiB).

Fuentes de licencia / License sources (checked 2026-11-06)

ReaZIT API metadata: author SassyAI; allowDerivatives=true, allowDifferentLicense=true, allowNoCredit=true, allowCommercialUse=["RentCivit","Image","Rent"]. These flags do not constitute unrestricted permission to sell model weights. The model description says not to sell or paywall the model. This converted release additionally follows the non-commercial scope communicated for this publication and always credits SassyAI.
Z-Image Turbo model card: Apache-2.0.
Qwen3-4B license: Apache-2.0.
Comfy-Org Z-Image Turbo repack: Apache-2.0; its VAE has the same SHA-256 as this ae.safetensors.

Nota / Note: model ID 779892 is an unrelated Marie Rose LoRA by Primarch, not ReaZIT. Verified ReaZIT ID: 2391667.

ReaZIT-QNN — paquete hermano / sibling package

La conversión histórica QNN/HTP v81 (SM8850) está en JiangFeng12/ReaZIT-QNN: paquete CL de la app QNN 0.25.0–0.25.2 y ZIP SL anterior de la línea 0.24.x. Son contextos QNN distintos al DiT FP8 NPH de este repo: no se usan con 0.30.1 NPH. La línea QNN tiene fallos conocidos y está congelada; consulta la tarjeta antes de descargar.

The historical QNN/HTP v81 (SM8850) conversion is available separately at JiangFeng12/ReaZIT-QNN: CL package for the 0.25.0–0.25.2 QNN app and older 0.24.x-era SL ZIP. These QNN contexts are not for the 0.30.1 NPH app. The QNN research line is frozen with known issues; read its card first.

ReaZIT original: SassyAI — modelo / model. Ambas distribuciones convertidas son no comerciales y conservan los términos originales y del modelo base. Both converted distributions are non-commercial and retain original/base-model terms.

Conversion tools and clearly marked herramientas teóricas, no probadas / theoretical, untested Krea 2 / Hunyuan research notes are linked from the QNN card. No models are included in those tools; existing code-license splits are retained.

Downloads last month: 13

GGUF

Model size

4B params

Architecture

qwen3

Hardware compatibility

8-bit

Model tree for JiangFeng12/ReaZIT-NPH

Base model

Tongyi-MAI/Z-Image-Turbo

Quantized

(106)

this model