Instructions to use lhca521/GLM-5-abliterated with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use lhca521/GLM-5-abliterated with Transformers:

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

pipe = pipeline("text-generation", model="lhca521/GLM-5-abliterated")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForMultimodalLM

tokenizer = AutoTokenizer.from_pretrained("lhca521/GLM-5-abliterated")
model = AutoModelForMultimodalLM.from_pretrained("lhca521/GLM-5-abliterated")
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

vLLM

How to use lhca521/GLM-5-abliterated with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "lhca521/GLM-5-abliterated"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "lhca521/GLM-5-abliterated",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/lhca521/GLM-5-abliterated

SGLang

How to use lhca521/GLM-5-abliterated 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 "lhca521/GLM-5-abliterated" \
    --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": "lhca521/GLM-5-abliterated",
		"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 "lhca521/GLM-5-abliterated" \
        --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": "lhca521/GLM-5-abliterated",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use lhca521/GLM-5-abliterated with Docker Model Runner:
```
docker model run hf.co/lhca521/GLM-5-abliterated
```

GLM-5-abliterated

File size: 1,339 Bytes

529c03e

---
license: apache-2.0
base_model: zai-org/GLM-5
tags:
  - abliterated
  - uncensored
  - glm
  - moe
library_name: transformers
---

# GLM-5 Abliterated (BF16)

""""""wont recommend using this, please let me know if u do"""""   . This is an abliterated  version of [zai-org/GLM-5](https://huggingface.co/zai-org/GLM-5) (744B MoE, 40B active parameters).

## What is abliteration?

Abliteration removes the "refusal direction" from the model weights using weight orthogonalization. This allows the model to respond to a wider range of prompts without safety refusals, while preserving general capability.

## Method

1. Computed refusal directions for all 78 layers using contrastive activation pairs (harmful vs harmless prompts)
2. Applied weight orthogonalization to layers 15-54:
   - `self_attn.o_proj.weight` (attention output projection)
   - `mlp.shared_experts.down_proj.weight` (shared expert down projection)
3. Alpha = 1.0, 80 weight matrices modified total

## Details

- **Base model**: zai-org/GLM-5 (744B MoE, BF16)
- **Modified layers**: 15-54 (40 of 78 total layers)
- **Weights modified**: 80 (o_proj + shared_experts.down_proj per layer)
- **Precision**: BF16 (full precision, no quantization artifacts)

## Disclaimer

This model is provided for research purposes. Users are responsible for ensuring appropriate use.