Instructions to use wardner/Qwen-Image-2.1 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Notebooks
Google Colab
Kaggle
Local Apps Settings

How to use wardner/Qwen-Image-2.1 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 wardner/Qwen-Image-2.1
# Run inference directly in the terminal:
llama cli -hf wardner/Qwen-Image-2.1

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf wardner/Qwen-Image-2.1
# Run inference directly in the terminal:
llama cli -hf wardner/Qwen-Image-2.1

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 wardner/Qwen-Image-2.1
# Run inference directly in the terminal:
./build/bin/llama cli -hf wardner/Qwen-Image-2.1

Use Docker

docker model run hf.co/wardner/Qwen-Image-2.1

Llama new
LM Studio
Jan
Ollama
How to use wardner/Qwen-Image-2.1 with Ollama:
```
ollama run hf.co/wardner/Qwen-Image-2.1
```
Unsloth Desktop

How to use wardner/Qwen-Image-2.1 with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf wardner/Qwen-Image-2.1

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": "wardner/Qwen-Image-2.1"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use wardner/Qwen-Image-2.1 with Docker Model Runner:
```
docker model run hf.co/wardner/Qwen-Image-2.1
```

Lemonade

How to use wardner/Qwen-Image-2.1 with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull wardner/Qwen-Image-2.1

Run and chat with the model

lemonade run user.Qwen-Image-2.1-{{QUANT_TAG}}

List all available models

lemonade list

Hermes Agent

How to use wardner/Qwen-Image-2.1 with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf wardner/Qwen-Image-2.1

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 wardner/Qwen-Image-2.1

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use wardner/Qwen-Image-2.1 with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf wardner/Qwen-Image-2.1

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 "wardner/Qwen-Image-2.1" \
  --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"

Qwen-Image-2.1 — [GEN]图片生成v2.1

English | 简体中文

1. Overview

Qwen-Image-2.1 text-to-image model: a prompt goes in, a 512×512 image comes out.

This repository ships the RK182x deployment artifacts (RKNN graphs + weights) together with the wdnwebui worker manifest (qwenimage21.json).

The pipeline is split across three compute devices:

Stage	Device	Runtime
Text encoder (TE) + DiT head	EP card 0 (RK182x)	RKNN3
DiT, 4 segments	EP card 1 (RK182x)	RKNN3
VAE decode (and VAE encode, for image-to-image)	the host RK3588's own NPU	RKNN2

Because the VAE is physically separate from the DiT, everything stays resident: the generation loop performs zero model_init. A second image costs only the sampling steps plus the decode.

Item	Value
Model ID	`wardner/Qwen-Image-2.1`
Display name	Qwen-Image-2.1
Model type	T2I (text-to-image)
Compute target	RK182x (×2) + host RK3588 NPU
Host platform	RK3588
Runtime	RKNN3 + RKNN2
Capabilities	`swcap_text_to_image`
Worker factory	`qwen_image_21_512`
Memory estimate	~10200 MB
Init time	~110 s
Cards required	2 EP cards (+ 1 local NPU for the VAE)
Total weight size	10.1 GB
Output	512×512 PNG

2. Files

Text encoder (RK182x)

Path	Description	Size
`llm_qwenimage21-te.1828.rknn`	RKNN graph (RK182x compute card)	7.0 MB
`llm_qwenimage21-te.weight`	Weights paired with the RKNN graph above	3.7 GB
`qwenimage21-te.embed.bin`	fp16 embedding table (vocab × hidden)	1.2 GB
`qwenimage21-te.tokenizer.gguf`	Tokenizer (GGUF, GPT-2 BPE + ByteLevel)	5.7 MB

DiT — head + 4 segments (RK182x)

Path	Description	Size
`dit_qwenimage21-512.head.1828.rknn`	RKNN graph — the DiT head	1.2 MB
`dit_qwenimage21-512.head.weight`	Weights paired with the RKNN graph above	46.9 MB
`dit_qwenimage21-512hc.seg0.1828.rknn`	RKNN graph — DiT segment 0/4	13.7 MB
`dit_qwenimage21-512hc.seg0.weight`	Weights paired with the RKNN graph above	989.0 MB
`dit_qwenimage21-512hc.seg1.1828.rknn`	RKNN graph — DiT segment 1/4	13.7 MB
`dit_qwenimage21-512hc.seg1.weight`	Weights paired with the RKNN graph above	970.3 MB
`dit_qwenimage21-512hc.seg2.1828.rknn`	RKNN graph — DiT segment 2/4	13.7 MB
`dit_qwenimage21-512hc.seg2.weight`	Weights paired with the RKNN graph above	970.3 MB
`dit_qwenimage21-512hc.seg3.1828.rknn`	RKNN graph — DiT segment 3/4	13.0 MB
`dit_qwenimage21-512hc.seg3.weight`	Weights paired with the RKNN graph above	979.7 MB
`dit_qwenimage21-512hc.int_ledger`	External per-core INT budget ledger (plain text)	40 B

VAE

Two mutually exclusive layouts ship in this package; the runtime picks the one the board can satisfy (see §6).

Path	Description	Size
`vae_qwenimage21-512.3588.rknn`	VAE decoder for the host RK3588 NPU (RKNN2)	581.6 MB
`vae_qwenimage21-512.encoder.3588.rknn`	VAE encoder for the host RK3588 NPU — image-to-image	205.1 MB
`vae_qwenimage21-512.seg0.1828.rknn`	VAE decoder, segment 0/4 (RK182x, RKNN3)	4.0 MB
`vae_qwenimage21-512.seg0.weight`	Weights paired with the RKNN graph above	391.6 MB
`vae_qwenimage21-512.seg1.1828.rknn`	VAE decoder, segment 1/4 (RK182x, RKNN3)	3.0 MB
`vae_qwenimage21-512.seg1.weight`	Weights paired with the RKNN graph above	71.3 MB
`vae_qwenimage21-512.seg2.1828.rknn`	VAE decoder, segment 2/4 (RK182x, RKNN3)	4.4 MB
`vae_qwenimage21-512.seg2.weight`	Weights paired with the RKNN graph above	12.7 MB
`vae_qwenimage21-512.seg3.1828.rknn`	VAE decoder, segment 3/4 (RK182x, RKNN3)	1.6 MB
`vae_qwenimage21-512.seg3.weight`	Weights paired with the RKNN graph above	1.6 MB

Release files

Path	Description	Size
`qwenimage21.json`	worker manifest — deliberately plain text	5.9 KB
`config.json`	Release metadata and file manifest	—
`SHA256SUMS`	SHA-256 checksums of every model file	—
`docs/README.md`	The model's own technical README	56.2 KB
`NOTICE`, `LICENSE`	Attribution and the upstream agreement	—

qwenimage21.json is deliberately not in Git LFS. It is the manual of this distribution unit and must stay plain text so it can be diffed, reviewed and previewed on the Hub.

3. Download

hf download wardner/Qwen-Image-2.1 --local-dir ./Qwen-Image-2.1

Or with git (requires Git LFS):

git clone https://huggingface.co/wardner/Qwen-Image-2.1

If huggingface.co is unreachable from your network (DNS interception / blocked access), point the client at a mirror instead:

HF_ENDPOINT=https://hf-mirror.com hf download wardner/Qwen-Image-2.1 \
  --local-dir ./Qwen-Image-2.1

Note: with git clone the weights are Git LFS pointers until git lfs pull completes. hf download fetches the real files directly, and does not create a .git directory — neither is required for the model to work; see §4.

4. Installation (wdnwebui)

Clone the repository into the model root so the folder holds both the weights and the worker manifest:

git clone https://huggingface.co/wardner/Qwen-Image-2.1 /userdata/models/Qwen-Image-2.1

Then open the web UI → Compute → the card appears under "local compute" (the backend rescans the model root every ~2 s, no restart needed) → click "enable".

The directory name does not have to match the model name: the SDK follows "the folder the manifest was found in", located through the WDNAI_QWEN_IMAGE_21_DIR environment variable declared in the manifest.

5. SHA-256 verification

cd ./Qwen-Image-2.1
sha256sum -c SHA256SUMS

All entries must report OK before deployment.

6. Caveats and limits

Two legal topologies, declared by which files are in the package:
- A — 2×RK182x + host RK3588: the package contains vae_qwenimage21-512.3588.rknn, so the VAE runs on the host's own RKNN2 NPU.
- B — 3×RK182x: the package contains vae_qwenimage21-512.seg{0..3}.1828.rknn, so the VAE (RKNN3) takes the third card exclusively. With neither present, the model fails fast at open time instead of silently falling back to a slow shared-card layout.
Cards are picked at runtime. The EP cards are enumerated from the device list (PCI BDF, ascending); if enumeration yields nothing and no card was given explicitly, the run errors out — no hard-coded address is ever used.
The VAE must stay fp16. A plain w8a8 quantization destroys the output (it collapses to a flat colour). The shipped VAE is auto-hybrid; see docs/README.md for the measured accuracy figures.
Sampling is CPU-bound at the tail. A few large Transposes fall back to the CPU, so the wall clock is dominated by those rather than by NPU throughput — int8 is not faster than fp16 here.
See the notes field of qwenimage21.json — it records the per-model specifics (topology, quantization, timings).

7. License

The upstream model is released under the Qwen RESEARCH LICENSE AGREEMENT — research and evaluation use only; commercial use requires a separate license from Hangzhou Tongyi Laboratory. The files here are RKNN-converted derivatives, not unmodified upstream weights.

Redistribution requires the upstream agreement to be shipped alongside (this is LICENSE) together with the attribution notice in NOTICE.

Downloads last month: 109

GGUF

Model size

0 params

Architecture

qwen3vl

Hardware compatibility

We're not able to determine the quantization variants.

View all variants