Instructions to use JiangFeng12/ReaZIT-QNN 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-QNN 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-QNN:Q8_0
# Run inference directly in the terminal:
llama cli -hf JiangFeng12/ReaZIT-QNN: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-QNN:Q8_0
# Run inference directly in the terminal:
llama cli -hf JiangFeng12/ReaZIT-QNN: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-QNN:Q8_0
# Run inference directly in the terminal:
./build/bin/llama cli -hf JiangFeng12/ReaZIT-QNN:Q8_0

Use Docker

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

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

How to use JiangFeng12/ReaZIT-QNN 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-QNN: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-QNN:Q8_0"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

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

Lemonade

How to use JiangFeng12/ReaZIT-QNN with Lemonade:

Pull the model

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

Run and chat with the model

lemonade run user.ReaZIT-QNN-Q8_0

List all available models

lemonade list

Hermes Agent

How to use JiangFeng12/ReaZIT-QNN 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-QNN: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-QNN:Q8_0

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use JiangFeng12/ReaZIT-QNN 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-QNN: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-QNN: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-QNN · Hexagon HTP v81 / SM8850

Español (México)

Conversión QNN/HTP de ReaZIT v1.0, realizada por JiangFeng12 para la línea histórica ZIT-HEXAGON 0.24.x / 0.25.x, con destino SM8850 / Hexagon v81 / QAIRT 2.50. Autora original: SassyAI — ReaZIT original. Gracias a SassyAI por permitir publicar la conversión con crédito y enlace a su perfil; la autorización fue comunicada por JiangFeng12. Esta no es una publicación oficial de SassyAI ni de Qualcomm.

Distribución convertida solo para uso no comercial. No MIT para los pesos. Aplican los términos originales de ReaZIT, los de Z-Image Turbo y las licencias por componente. No vender los pesos ni ponerlos detrás de un muro de pago. Conserva el crédito y los enlaces de SassyAI.

Dos paquetes diferentes: no mezclar

ReaZIT-0.25/ (CL): carpeta completa para la ruta QNN de ZIT-HEXAGON 0.25.0–0.25.2. zit.json declara minApp=0.25.0, formatVersion=2, LoRA swap rango 96 (hasta 7, recomendadas 3). Seis bloques DiT con pesos INT8 por canal, activaciones/E/S FP16; VAE FP16; TE Qwen3-4B Q8_0 en CPU. t_embedder.bin es un pequeño MLP FP32 para el host, no un runtime del SDK.
legacy/ReaZIT-npu-v81.zip (SL): paquete anterior, 0.24.x, con carpeta interna ReaZIT/, formatVersion=2, mismo destino QNN 2.50/v81, sin entradas LoRA (supported=false). Sus DiT tienen bytes distintos a los del paquete 0.25. No sustituye al paquete CL.
Resolución fija 1024×1024, valores iniciales: Euler / simple, 8 pasos, CFG 1.

Usa la carpeta 0.25 como modelo nuevo (sin reemplazar la anterior), verifica SHA256SUMS desde la raíz del repositorio y selecciónala en la app QNN histórica. Para el ZIP SL usa la importación ZIP de la app histórica. Se necesita espacio para ZIP + extracción.

Archivo de investigación, no una recomendación de calidad: la línea QNN está congelada; el informe documenta NaN de caption tokens en bloque 0 en el dispositivo, imágenes de cuadrícula y límites de RAM. Que un paquete sea compatible con el formato no significa que esos fallos se hayan resuelto.

No son los pesos NPH de 0.30.1. Desde 0.26 la app cambió de motor; para la 0.30.1 recomendada usa el repo hermano JiangFeng12/ReaZIT-NPH y el GitHub NPH.

English

QNN/HTP conversion of ReaZIT v1.0 by JiangFeng12 for the historical ZIT-HEXAGON 0.24.x / 0.25.x QNN path, targeting SM8850 / Hexagon HTP v81 / QAIRT 2.50. Original author: SassyAI — original ReaZIT model. Thank you to SassyAI for permitting conversion publication with author credit and a profile link. Permission was reported by JiangFeng12; this is not an official SassyAI or Qualcomm release.

Non-commercial distribution only; weights are not MIT. The original ReaZIT terms, Z-Image Turbo base-model terms and per-component licenses also apply. Do not sell/paywall the weights. Keep SassyAI's credit and links.

ReaZIT-0.25/ (CL) works with the historical 0.25.0–0.25.2 QNN app path. Manifest: minApp=0.25.0, format version 2, rank-96 swap LoRAs (up to 7, 3 recommended), six INT8-weight/FP16-activation DiT contexts, FP16 VAE, CPU Qwen3-4B Q8_0 text encoder and FP32 host time embedder.
legacy/ReaZIT-npu-v81.zip (SL) is the older 0.24.x-era package, contains its own ReaZIT/zit.json, and has no LoRA inputs. Its DiT files are different; do not mix package files.
Fixed 1024×1024; initial settings Euler/simple, 8 steps, CFG 1. Import the full 0.25 folder as a new model or import the legacy ZIP in the historical QNN app. Verify SHA256SUMS; ZIP import needs space for both archive and extracted files.

Research archive, not a quality recommendation: the archived QNN project records unresolved on-device block-0 caption NaNs/grid images and RAM/residency limits. Format compatibility is not a claim these bugs are fixed.

For the separate 0.30.1 NPH app, use ReaZIT-NPH and Xiemox2/zit-hexagon-nph, not these contexts.

Herramientas / Tools

ReaZIT QNN — código propio MIT / original MIT code

The matching public tools repository already exists; tools are linked rather than duplicated:

Xiemox2/ZIT-HEXAGON-QNN, verified revision 6002fd59aa93dfa05650eecdb80b45358136ef35.
Conversion scripts: py/export_chunk.py, export_chunk_lora.py, zit_enc.py, export_vae.py, ESRGAN/TE/reference/validation scripts; sh/run_chunk.sh, run_all_chunks.sh, post_chunks.sh, conv_te.sh, run_vae.sh, run_esrgan.sh, assemble_pkg.sh; cfg/ HTP settings.
Op dictionary and node encodings, including zit_op_dictionary.json.
Conversion/build README.

The source archive's older statements that model permission is pending are historical; the present conversion release follows the later permission reported by JiangFeng12. Code's MIT license does not relicense weights.

Krea 2 / Hunyuan QNN — herramientas teóricas, no probadas / theoretical, untested

El usuario no continuará estas líneas por ahora; se dejan para que otras personas las terminen. / The user will not continue these lines for now; others are welcome to finish them.

Krea 2 QNN: existing public Xiemox2/Tools_Krea2-npu, verified revision 223dc454712b32a8051d9994476f18c467abb376. Theoretical, untested on device; no real-model QNN context generation or phone execution was completed. Host/toy checks do not constitute hardware validation. No model weights included. Its existing code license is Apache-2.0, not MIT; retain its LICENSE/NOTICE and do not relicense derived Apache code. Krea weight terms are separate.
Hunyuan QNN: Xiemox2/Tools_Hunyuan-qnn — herramientas teóricas, no probadas / theoretical, untested (HunyuanVideo 1.5 y 13B → HTP v81; plan de segmentos, export ONNX, qairt-converter/quantizer, context binaries, manifiesto hv.json). Los adaptadores del DiT y del VAE fallan a propósito hasta implementarse / DiT and VAE adapters fail on purpose until implemented. MIT, sin pesos / no weights. La app Hunyuan funcional usa ggml-hexagon, no QNN: hunyuan-npu-models.

Qualcomm / Real-ESRGAN license review

The model packages contain the user's generated conversion outputs, not a standalone Qualcomm SDK/runtime mirror. The separately provided historical APKs under app/ do bundle Qualcomm runtime .so libraries and an HTP v81 skel; see the application section below. No SDK headers or compiler tools are provided. The legacy ZIP directory was inspected remotely: model binaries, GGUF, manifest and the two notices only; no SDK libraries/executables. No test/ binaries or paid SDXL assets are copied.

The supplied QNN_NOTICE.txt is a third-party attribution notice, not the SDK EULA. Its opening says:

your use of such software components together with the QTI software ("Software") is subject to the terms of your separate license from QTI.

Except as may be granted by separate express written agreement, this file provides no license to any patents, trademarks, copyrights, or other intellectual property of QTI or its affiliates.

No clause clearly prohibiting redistribution of the user's generated model outputs was identified in that notice. Publication proceeds on the user's explicit instruction that these are their own conversion outputs, not on a claim that the notice grants SDK redistribution rights. The installed SDK attribution files and public SDK legal documentation were checked; the applicable full EULA was not available in the local installation and the public archive fetch returned HTTP 403. Thus this is not a definitive interpretation or certification of Qualcomm's separate EULA. Obtain/use QAIRT runtime under your own Qualcomm agreement; no Qualcomm runtime rights are granted here.

Z-Image Turbo and Qwen3-4B retain Apache-2.0 terms. Original ReaZIT terms: Civitai. ReaZIT conversion distribution is non-commercial; code and model licenses are distinct.

Archivos copiados / Copied files: sizes and SHA-256

Source/destination LFS SHA-256 object identifiers were compared; small text files were hashed locally. The model weights were copied server-side, not downloaded locally. The top-level SHA256SUMS covers every copied source file, including the original nested checksum manifest.

Path	Bytes	SHA-256
`ReaZIT-0.25/LEEME.md`	938	`0a62d508779480a5d4f1215e9953a45e2999f2fb3a45c20e36ad9675a764f95f`
`ReaZIT-0.25/SHA256SUMS`	1,065	`73935f153d6a02d98e580a1488bdf03635c28a2972307e34f8399725e894cc1e`
`ReaZIT-0.25/dit_0.bin`	844,974,872	`1bb9597ae09934a950ce3a483f519ac2c9f6ef29239c74ec2bf30eff404d08fe`
`ReaZIT-0.25/dit_1.bin`	1,426,304,056	`34d7e8be3f997bac61faac29e70a8b2794cb14bee457c12d19bf8211222c58ae`
`ReaZIT-0.25/dit_2.bin`	1,428,888,576	`2c551dc4b08b4260982426feaf6b8f9ae807c70f74b021e8b341f335a8b8cb2f`
`ReaZIT-0.25/dit_3.bin`	1,427,311,744	`ff3c5fc5d2192314e4be4cecad5326afd9a03bcec5b7661cd6ae4dac772863e2`
`ReaZIT-0.25/dit_4.bin`	1,429,871,576	`a7cb97e74399cf8be816386dca206431e51ef54272314d25f06b64fddc84647f`
`ReaZIT-0.25/dit_5.bin`	1,429,491,072	`2edeb5855a45f3a33b100fb2df17f85b462d32b56e931741401f4a021249cfcb`
`ReaZIT-0.25/licenses/QNN_NOTICE.txt`	130,445	`0c5e8aad3506d0ab881cabaf0dae8de64e1784dc1ee6c873caf24a78fbf01924`
`ReaZIT-0.25/licenses/Real-ESRGAN_LICENSE.txt`	1,519	`4a699ec4863d96a91fc265948a0c90033f7e8735d515524dcf3444736406e0c2`
`ReaZIT-0.25/realesrgan_x4plus_t256.bin`	52,611,256	`70f98bf551296a05c72f16c61324919f9baacfa701935cfdcce1f80def9f3c85`
`ReaZIT-0.25/t_embedder.bin`	2,102,272	`b64e149531a2158cc5598aec2d3951b7a988c6617e0e69771cbb85e3f4d36d19`
`ReaZIT-0.25/text_encoder_q8_0.gguf`	4,280,404,928	`0a2cee0a56ce2c8499f9ac607a10b9ef06a4e71fbe0a781ffa5b2f54c5d9ab8e`
`ReaZIT-0.25/vae_decoder.bin`	169,851,464	`161f227fa3ca4c097ecf266882a2b059282834d52a142ecf40f0a9bd48b427d4`
`ReaZIT-0.25/zit.json`	1,576	`2556312deb80bb85f402f8f7ea7657aa96fe502054944094b9cd967ee8957a6a`
`legacy/ReaZIT-npu-v81.zip`	12,400,227,867	`4e583f88dcaaa64660d7f87f3434a4db43dc6a61f2b0a07583168dc7d352d425`

Total copied source files: 24,892,175,226 bytes (~23.18 GiB), including both variants; do not download both unless needed. The checksum table documents the archive file itself, not individually extracted legacy members.

App ZIT-HEXAGON QNN histórica / Historical QNN app

Código fuente / source: Xiemox2/ZIT-HEXAGON-QNN. Paquete Android / Android package: com.zithexagon (arm64).

APK	versionCode	Bytes	Paquete compatible / Compatible model package	SHA-256
`zit-hexagon-0.24.3-release.apk`	29	19,777,062	Legacy SL ZIP (0.24.x)	`79840c15227a624798eed7a6c210d9aee60a4b709bb9a1cf08c306cfb48dd655`
`zit-hexagon-0.25.0-release.apk`	31	19,929,910	ReaZIT-0.25 CL folder	`917a0ced0f3134134a35e2836b7a42ad2a4d8a6cccae08e106408ec214748f77`
`zit-hexagon-0.25.1-release.apk`	32	19,965,722	ReaZIT-0.25 CL folder	`6715848bc86c45d485e46418ab7174c1f6b206ca857b8b0c87e744c34163fee3`
`zit-hexagon-0.25.2-release.apk`	33	19,996,446	ReaZIT-0.25 CL folder	`01e1d1bc377dbd3c6e2a39af37912da6c701942650ccbe9ee8e3f8dd2619d4f2`

Español (México): para la carpeta ReaZIT-0.25/, utiliza la app QNN 0.25.0–0.25.2; la 0.25.2 es la última versión archivada de esa ruta. La 0.24.3 corresponde al ZIP SL anterior sin LoRA; no carga el paquete CL como si fuera el mismo modelo. Son APKs históricos de investigación con los fallos QNN descritos arriba, no la app NPH 0.30.1 recomendada. Verifica los SHA-256 antes de instalar; no se probó la ejecución de estas APKs en esta publicación.

English: use the historical QNN app 0.25.0–0.25.2 for ReaZIT-0.25/; 0.25.2 is the last archived build of that path. Version 0.24.3 matches the older SL ZIP without LoRA; the CL package is not interchangeable. These research APKs retain the known QNN issues above and are not the recommended 0.30.1 NPH app. Verify SHA-256 before installation; no APK was executed as part of this publication.

Qualcomm runtime in the APKs / runtime de Qualcomm dentro de las APKs

All four APKs bundle these proprietary Qualcomm files under lib/arm64-v8a/:

libQnnHtp.so
libQnnHtpV81.so
libQnnHtpV81Skel.so
libQnnHtpV81Stub.so
libQnnSystem.so

Aviso EULA / EULA caveat: la licencia MIT del código propio no licencia estas bibliotecas. Su uso y redistribución están sujetos al acuerdo independiente aplicable de Qualcomm QAIRT/QNN. QNN_NOTICE.txt es un aviso de terceros, no una licencia suficiente para redistribuir el SDK. No se ha establecido una autorización definitiva mediante el EULA completo; consulta y cumple tu acuerdo de Qualcomm. Esta publicación no concede derechos de Qualcomm ni afirma autorización expresa de Qualcomm.

The MIT license of the app's original code does not license these libraries. Use and redistribution remain subject to the applicable separate Qualcomm QAIRT/QNN agreement. The bundled QNN third-party notice is not itself an SDK redistribution license. Definitive clearance under the full applicable EULA has not been established; consult and comply with your Qualcomm agreement. This release grants no Qualcomm rights and claims no express Qualcomm authorization. These are complete historical applications, not a standalone SDK/runtime mirror.

Revisión estática / Static inspection

The four APKs were inspected as ZIP archives and Android manifests. VersionCodes above were extracted with aapt2 dump badging; SHA-256 matched the private source objects. No bundled model checkpoints (.safetensors, .gguf, .onnx, .dlc), keystores or paid SDXL model assets were found; the only asset over 100 KB is the QNN attribution notice. Exact publication-token, standalone token/private-key-pattern and asset-name checks found no secrets. Raw hf_ substrings inside two Qualcomm binaries were verified as byte-identical to the installed official SDK files, not added credentials. This static review is not a complete security audit or a hardware compatibility test.

Downloads last month: 98

GGUF

Model size

4B params

Architecture

qwen3

Hardware compatibility

8-bit

Model tree for JiangFeng12/ReaZIT-QNN

Base model

Tongyi-MAI/Z-Image-Turbo

Quantized

(107)

this model