Instructions to use pottokao/Z-Image-Engineer-V6-DFlash-b16 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Notebooks
Google Colab
Kaggle
Local Apps Settings

How to use pottokao/Z-Image-Engineer-V6-DFlash-b16 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 pottokao/Z-Image-Engineer-V6-DFlash-b16:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf pottokao/Z-Image-Engineer-V6-DFlash-b16:Q4_K_M

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf pottokao/Z-Image-Engineer-V6-DFlash-b16:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf pottokao/Z-Image-Engineer-V6-DFlash-b16:Q4_K_M

Use pre-built binary

# Download pre-built binary from:
# https://github.com/ggerganov/llama.cpp/releases
# Start a local OpenAI-compatible server with a web UI:
./llama-server -hf pottokao/Z-Image-Engineer-V6-DFlash-b16:Q4_K_M
# Run inference directly in the terminal:
./llama-cli -hf pottokao/Z-Image-Engineer-V6-DFlash-b16:Q4_K_M

Build from source code

git clone https://github.com/ggerganov/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-server llama-cli
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama-server -hf pottokao/Z-Image-Engineer-V6-DFlash-b16:Q4_K_M
# Run inference directly in the terminal:
./build/bin/llama-cli -hf pottokao/Z-Image-Engineer-V6-DFlash-b16:Q4_K_M

Use Docker

docker model run hf.co/pottokao/Z-Image-Engineer-V6-DFlash-b16:Q4_K_M

LM Studio
Jan
Ollama
How to use pottokao/Z-Image-Engineer-V6-DFlash-b16 with Ollama:
```
ollama run hf.co/pottokao/Z-Image-Engineer-V6-DFlash-b16:Q4_K_M
```
Unsloth Desktop

How to use pottokao/Z-Image-Engineer-V6-DFlash-b16 with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf pottokao/Z-Image-Engineer-V6-DFlash-b16:Q4_K_M

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": "pottokao/Z-Image-Engineer-V6-DFlash-b16:Q4_K_M"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use pottokao/Z-Image-Engineer-V6-DFlash-b16 with Docker Model Runner:
```
docker model run hf.co/pottokao/Z-Image-Engineer-V6-DFlash-b16:Q4_K_M
```

Lemonade

How to use pottokao/Z-Image-Engineer-V6-DFlash-b16 with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull pottokao/Z-Image-Engineer-V6-DFlash-b16:Q4_K_M

Run and chat with the model

lemonade run user.Z-Image-Engineer-V6-DFlash-b16-Q4_K_M

List all available models

lemonade list

Hermes Agent

How to use pottokao/Z-Image-Engineer-V6-DFlash-b16 with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf pottokao/Z-Image-Engineer-V6-DFlash-b16:Q4_K_M

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 pottokao/Z-Image-Engineer-V6-DFlash-b16:Q4_K_M

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use pottokao/Z-Image-Engineer-V6-DFlash-b16 with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf pottokao/Z-Image-Engineer-V6-DFlash-b16:Q4_K_M

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 "pottokao/Z-Image-Engineer-V6-DFlash-b16:Q4_K_M" \
  --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"

Z-Image-Engineer-V6 DFlash draft (b16)

A DFlash block-diffusion draft head for Z-Image-Engineer V6 with the Master Style prompt-rewrite LoRA attached (pottokao/zeng-v6_master-pe_lora_v4.1-r32-ep2_20260928), i.e. the Master Style role-play system prompt + style card workload on the rewrite (PE) role of the ComfyUI-DualHead-Dragon node. Started from z-lab/Qwen3-4B-DFlash-b16; same architecture and tensor names (5 layers, block 16, target layers 1/9/17/25/33).

It is specialised: it only speeds up that workload. On other prompts it still gives correct output (speculative decoding never changes what the target produces) but guesses less often.

Files

file	MB	use
`model.safetensors` + `config.json` + `dflash.py`…	1075	z-lab / transformers / vLLM format
`gguf/z-image-engineer-v6-dflash-b16-bf16.gguf`	1081	llama.cpp reference
`gguf/z-image-engineer-v6-dflash-b16-Q8_0.gguf`	577	near-lossless
`gguf/z-image-engineer-v6-dflash-b16-Q6_K.gguf`	447	recommended
`gguf/z-image-engineer-v6-dflash-b16-Q4_K_M.gguf`	323	smallest; same speed as Q6_K

The GGUFs were converted with convert_hf_to_gguf.py --target-model-dir <V6 HF> and quantized with llama-quantize from the bf16 file. The draft shares the target's token embedding and output head, so the files hold only the 5 draft layers.

llama-server -m <V6 gguf> --lora <master PE LoRA gguf> -md gguf/z-image-engineer-v6-dflash-b16-Q6_K.gguf \
  --spec-type draft-dflash --spec-draft-n-max 15 -fa on -np 1

Single sequence only: upstream draft-dflash acceptance collapses with many parallel sequences (llama.cpp #27117). --spec-draft-n-max 15: llama.cpp verifies 9–16 tokens at almost the cost of one, 5–8 is the expensive range.

Results

Offline, 120 held-out prompts, K=15:

	z-lab original	this head
tau (tokens advanced per verify, incl. bonus)	2.096	3.977
position 1..8 accuracy	.58 .40 .29 .22 .18 .15 .14 .12	.78 .65 .54 .46 .40 .35 .31 .28

llama.cpp (native draft-dflash, K=15), target V6 DHQ 2.9 GB + master PE LoRA v4.1 (f16), 10 Master Style prompts with different cards, greedy, 1000 new tokens max, one RTX 5060 Ti 16 GB:

draft	tok/s	speed-up	acceptance
none	87.0	1.00×	-
bf16	151.5	1.74×	16.8%
Q8_0	159.3	1.83×	16.7%
Q6_K	161.2	1.85×	16.7%
Q4_K_M	162.4	1.87×	16.6%

bf16 / Q8_0 / Q6_K give identical outputs on all 10 prompts, Q4_K_M on 9 of 10 (one greedy near-tie). With the LoRA merged into the target instead of attached at runtime: 107.4 → 209.0 tok/s (Q6_K draft, 2.40× over the LoRA-attached baseline).

Training

Data: 9,072 Master Style prompts regenerated with V6 + the master PE LoRA (v4) (vLLM, temp 0.7 / top_p 0.8 / top_k 20, thinking off); 326 held out.
Target features: V6 bf16 with the LoRA merged. anchors 128 / 2 epochs / lr 1e-4 cosine / 4 seqs per step / position-decay gamma 7 / 8-bit AdamW; 4,532 steps (train_dflash_v6.py, 2x RTX 5060 Ti, 4.56 s/step).

sha256

model.safetensors 52bee0ee736c70a3f672e7ab741606ae453b5fd74d607cd1dcd1dbcfad47ccb4
bf16 gguf d9a7aac5af73d7047751a7b3a40cdde6e9ce9d69cd5f7a8a4f61d4f70175cd9b
Q8_0 gguf 3b66ee8e7b29eeca3452ce57b6f1080f498332c5f5c990a08ae1e4d19d460cba
Q6_K gguf 251ffb5b431f3eb3af7a02721e5f544365ceebb39ca1a82ba9019b5a688072c2
Q4_K_M gguf 2a35df57a79be2f6a8c883d691f8581b61aa2ed994a191a84bc60514855ba7a8

License

CC BY-NC 4.0 (non-commercial, free); commercial use needs a separate license from the author. It covers only these weights: the base model and the image models (DiTs) keep their own licenses. Details and upstream notices: LICENSE.md.

Downloads last month: 672

Safetensors

Model size

0.5B params

Tensor type

BF16

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for pottokao/Z-Image-Engineer-V6-DFlash-b16

Base model

z-lab/Qwen3-4B-DFlash-b16

Quantized

(4)

this model