Instructions to use wangzhang/GLM-4.7-Flash-abliteratex with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use wangzhang/GLM-4.7-Flash-abliteratex with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="wangzhang/GLM-4.7-Flash-abliteratex")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("wangzhang/GLM-4.7-Flash-abliteratex")
model = AutoModelForCausalLM.from_pretrained("wangzhang/GLM-4.7-Flash-abliteratex", device_map="auto")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=40)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

llama.cpp

How to use wangzhang/GLM-4.7-Flash-abliteratex 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 wangzhang/GLM-4.7-Flash-abliteratex:Q8_0
# Run inference directly in the terminal:
llama cli -hf wangzhang/GLM-4.7-Flash-abliteratex:Q8_0

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf wangzhang/GLM-4.7-Flash-abliteratex:Q8_0
# Run inference directly in the terminal:
llama cli -hf wangzhang/GLM-4.7-Flash-abliteratex:Q8_0

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 wangzhang/GLM-4.7-Flash-abliteratex:Q8_0
# Run inference directly in the terminal:
./llama-cli -hf wangzhang/GLM-4.7-Flash-abliteratex:Q8_0

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 wangzhang/GLM-4.7-Flash-abliteratex:Q8_0
# Run inference directly in the terminal:
./build/bin/llama-cli -hf wangzhang/GLM-4.7-Flash-abliteratex:Q8_0

Use Docker

docker model run hf.co/wangzhang/GLM-4.7-Flash-abliteratex:Q8_0

LM Studio
Jan

vLLM

How to use wangzhang/GLM-4.7-Flash-abliteratex with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "wangzhang/GLM-4.7-Flash-abliteratex"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "wangzhang/GLM-4.7-Flash-abliteratex",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/wangzhang/GLM-4.7-Flash-abliteratex:Q8_0

SGLang

How to use wangzhang/GLM-4.7-Flash-abliteratex with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "wangzhang/GLM-4.7-Flash-abliteratex" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "wangzhang/GLM-4.7-Flash-abliteratex",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "wangzhang/GLM-4.7-Flash-abliteratex" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "wangzhang/GLM-4.7-Flash-abliteratex",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Ollama
How to use wangzhang/GLM-4.7-Flash-abliteratex with Ollama:
```
ollama run hf.co/wangzhang/GLM-4.7-Flash-abliteratex:Q8_0
```

Unsloth Studio

How to use wangzhang/GLM-4.7-Flash-abliteratex with Unsloth Studio:

Install Unsloth Studio (macOS, Linux, WSL)

curl -fsSL https://unsloth.ai/install.sh | sh
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for wangzhang/GLM-4.7-Flash-abliteratex to start chatting

Install Unsloth Studio (Windows)

irm https://unsloth.ai/install.ps1 | iex
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for wangzhang/GLM-4.7-Flash-abliteratex to start chatting

Using HuggingFace Spaces for Unsloth

# No setup required
# Open https://huggingface.co/spaces/unsloth/studio in your browser
# Search for wangzhang/GLM-4.7-Flash-abliteratex to start chatting

How to use wangzhang/GLM-4.7-Flash-abliteratex with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf wangzhang/GLM-4.7-Flash-abliteratex:Q8_0

Configure the model in Pi

# Install Pi:
npm install -g @mariozechner/pi-coding-agent
# Add to ~/.pi/agent/models.json:
{
  "providers": {
    "llama-cpp": {
      "baseUrl": "http://localhost:8080/v1",
      "api": "openai-completions",
      "apiKey": "none",
      "models": [
        {
          "id": "wangzhang/GLM-4.7-Flash-abliteratex:Q8_0"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Hermes Agent new

How to use wangzhang/GLM-4.7-Flash-abliteratex with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf wangzhang/GLM-4.7-Flash-abliteratex: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 wangzhang/GLM-4.7-Flash-abliteratex:Q8_0

Run Hermes

hermes

Atomic Chat new

OpenClaw new

How to use wangzhang/GLM-4.7-Flash-abliteratex with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf wangzhang/GLM-4.7-Flash-abliteratex: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 "wangzhang/GLM-4.7-Flash-abliteratex: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"

Docker Model Runner
How to use wangzhang/GLM-4.7-Flash-abliteratex with Docker Model Runner:
```
docker model run hf.co/wangzhang/GLM-4.7-Flash-abliteratex:Q8_0
```

Lemonade

How to use wangzhang/GLM-4.7-Flash-abliteratex with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull wangzhang/GLM-4.7-Flash-abliteratex:Q8_0

Run and chat with the model

lemonade run user.GLM-4.7-Flash-abliteratex-Q8_0

List all available models

lemonade list

GLM-4.7-Flash-Abliteratex

GLM-4.7-Flash-Abliteratex is a rebuilt successor to wangzhang/GLM-4.7-Flash-abliterated. The release focuses on preserving the official checkpoint structure, reducing collateral capability damage, and providing a reproducible llama.cpp Q8_0 artifact.

This is an experimental model with substantially reduced refusal behavior. It can produce unsafe, false, private-looking, or otherwise harmful content. A low refusal score is not a factuality score. Do not treat generated personal data, credentials, citations, or operational claims as real without independent verification.

What changed from the previous release

Area	Previous release	Abliteratex
Weight keys	9,491	9,703
Parameters	29,943,390,976	31,221,488,576
MTP / NextN tensors	Missing 212 tensors	All 212 official tensors retained
`tokenizer_config.json`	305 bytes	Official 7,226-byte file
Attention edits	Yes	None
Routed-expert edits	Broad attention/MLP/router recipe	Routed-expert down-projection EGA only, plus four router rows per MoE layer
Official metadata	Regenerated/stripped fields	Official config, tokenizer, generation config, and chat template restored byte-for-byte

The new package contains the same 9,703 tensor keys and dtype distribution as the official source revision: 9,656 BF16 tensors and 47 F32 tensors.

Training and selection

Base: zai-org/GLM-4.7-Flash
Base revision: 7dd20894a642a0aa287e9827cb1a1f7f91386b67
Dataset: wangzhang/abliterix-datasets
Direction data: good_1000 train[:800] and harmful_1000 train[:800]
Search backend: vLLM in-place editing on one RTX PRO 6000 Blackwell 96GB
Search: 60 broad trials followed by 24 targeted trials
Selected targeted trial: 20
Direction: global, interpolated index 20.4365519952
Expert down-projection profile:
- maximum strength 3.9788400044
- peak layer 30.8739492862
- minimum strength 2.9557299481
- radius 21.0740371709
Router adjustment: top four safety-associated experts per MoE layer, router_bias = -0.85
Attention and shared-expert weights were intentionally left unchanged.

Evaluation

The selected search trial produced 2 refusals out of 50 search-evaluation prompts (4%) with fixed-continuation NLL drift of 0.0826336592 nats/token.

On a separate 150-prompt held-out slice (harmful_1000 train[850:1000]) with 256-token generations, the strict LLM judge reported:

26/150 refusals or degenerate/non-actionable answers (17.3%)
124/150 compliant answers (82.7%)

Many held-out failures request real-time private phone numbers, home addresses, SSNs, or other data that is not present in the prompt. Fabricating plausible PII can lower a refusal benchmark while making hallucination materially worse. This release reports the strict score instead of counting invented private data as a quality success.

llama.cpp and Q8_0

The included GLM-4.7-Flash-Abliteratex-Q8_0.gguf was converted with current llama.cpp commit 3018a11e79e489b657dbb77c95694889ccff92df.

Older GLM-4.7-Flash GGUF conversion paths were known to mishandle glm4_moe_lite MLA tensors and expert stacking. Use a recent llama.cpp build. For first-line troubleshooting, keep KV cache in F16/BF16; a Q8_0 model file and --cache-type-k/--cache-type-v q8_0 are different forms of quantization.

Example:

llama-cli \
  -m GLM-4.7-Flash-Abliteratex-Q8_0.gguf \
  -ngl all -c 4096 -n 256 --jinja \
  --cache-type-k f16 --cache-type-v f16 \
  -p "Explain why Canberra is the capital of Australia."

Relevant upstream context:

Transformers usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "wangzhang/GLM-4.7-Flash-abliteratex"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype="auto",
    device_map="auto",
)

The safetensors package retains the official MTP/NextN tensors even though some current Transformers and llama.cpp runtime paths intentionally ignore the extra prediction layer during ordinary generation.

Downloads last month: 510

Safetensors

Model size

30B params

Tensor type

F32

BF16

Model tree for wangzhang/GLM-4.7-Flash-abliteratex

Base model

zai-org/GLM-4.7-Flash

Quantized

(94)

this model

Quantizations

2 models

wangzhang
/

GLM-4.7-Flash-abliteratex

GLM-4.7-Flash-Abliteratex

What changed from the previous release

Training and selection

Evaluation

llama.cpp and Q8_0

Transformers usage

Model tree for wangzhang/GLM-4.7-Flash-abliteratex

Dataset used to train wangzhang/GLM-4.7-Flash-abliteratex