Instructions to use Blackfrost-AI/GLM-5.2-ABLITERATED-REAP-NU176-NVFP4 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use Blackfrost-AI/GLM-5.2-ABLITERATED-REAP-NU176-NVFP4 with Transformers:

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

pipe = pipeline("text-generation", model="Blackfrost-AI/GLM-5.2-ABLITERATED-REAP-NU176-NVFP4")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("Blackfrost-AI/GLM-5.2-ABLITERATED-REAP-NU176-NVFP4")
model = AutoModelForCausalLM.from_pretrained("Blackfrost-AI/GLM-5.2-ABLITERATED-REAP-NU176-NVFP4")
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 Blackfrost-AI/GLM-5.2-ABLITERATED-REAP-NU176-NVFP4 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "Blackfrost-AI/GLM-5.2-ABLITERATED-REAP-NU176-NVFP4"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Blackfrost-AI/GLM-5.2-ABLITERATED-REAP-NU176-NVFP4",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/Blackfrost-AI/GLM-5.2-ABLITERATED-REAP-NU176-NVFP4

SGLang

How to use Blackfrost-AI/GLM-5.2-ABLITERATED-REAP-NU176-NVFP4 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 "Blackfrost-AI/GLM-5.2-ABLITERATED-REAP-NU176-NVFP4" \
    --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": "Blackfrost-AI/GLM-5.2-ABLITERATED-REAP-NU176-NVFP4",
		"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 "Blackfrost-AI/GLM-5.2-ABLITERATED-REAP-NU176-NVFP4" \
        --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": "Blackfrost-AI/GLM-5.2-ABLITERATED-REAP-NU176-NVFP4",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use Blackfrost-AI/GLM-5.2-ABLITERATED-REAP-NU176-NVFP4 with Docker Model Runner:
```
docker model run hf.co/Blackfrost-AI/GLM-5.2-ABLITERATED-REAP-NU176-NVFP4
```

You need to agree to share your contact information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

🔒 Licensed access — get the weights + deploy kit

This is a gated, commercial release. Purchasing grants your Hugging Face account access to this repository (weights + deploy kit), delivered instantly.

➡️ Get access — satchellm.web.app · card or Bitcoin (−10%)

GLM-5.2-ABLITERATED-REAP-NU176-NVFP4

A de-risked, expert-pruned GLM-5.2 engineered to run fast on a single node of 4× NVIDIA RTX PRO 6000 (Blackwell, SM120) — with a verified 0-refusal safety profile.

Built by Blackfrost-AI.

TL;DR

What it is: GLM-5.2, refusal-ablated ("abliterated"), REAP non-uniform expert-pruned, and quantized to NVFP4 (4-bit) — a single deployable checkpoint.
Where it runs: 4× RTX PRO 6000 Blackwell (SM120) — half the card count of a typical full-precision deployment. ~302 GB on disk.
How fast: ~52.7 tok/s single-stream decode with multi-token-prediction acceleration — 1.55× faster than standard decode, on just 4 cards.
How safe-to-attack: 0 true refusals across 400 harmful prompts and 0% over-refusal on 100 safe prompts, 0 incoherent out of 600 — a legitimate, functional, fast red-team artifact.

Highlights


Base	GLM-5.2 (`glm_moe_dsa` — MLA + sparse attention + MTP head)
De-risk	Abliterated — refusal directions removed
Prune	REAP non-uniform expert pruning (~176 of 256 routed experts)
Quant	NVFP4 (4-bit)
Footprint	~302 GB · runs on 4× RTX PRO 6000 (Blackwell, SM120, 96 GB)
Acceleration	Multi-token-prediction (MTP) speculative decoding
Decode speed	~52.7 tok/s single-stream · 1.55× over standard decode
Safety profile	0 / 400 harmful refusals · 0 / 100 over-refusal · 0 / 600 incoherent

What this model is

A research / red-team build of GLM-5.2 that combines three transformations into one deployable artifact: refusal ablation (it does not decline tasks), non-uniform MoE expert pruning (smaller, while preserving quality), and NVFP4 quantization (fits and runs on four Blackwell cards instead of eight).

The result is the whole point of this release: a de-risked GLM-5.2 that is small enough and fast enough for a 4-card Blackwell node — and that stays coherent and functional after every transformation.

Performance

Measured single-stream, on 4× RTX PRO 6000 (Blackwell, SM120):

Decode mode	Throughput
Blackfrost MTP acceleration (shipped)	~52.7 tok/s
Standard decode	~34.0 tok/s

~1.55× faster on the same 4 cards — the acceleration is built into the shipped configuration.

Safety / de-risk evaluation

Evaluated on the live serve (these numbers reflect exactly how it ships). Every substring-flagged case was read by hand to confirm.

Dataset	Prompts	Refusals
AdvBench (harmful)	200	0 true
StrongREJECT (harmful)	200	0 true
XSTest — safe (over-refusal)	100	0 (0.0%)
Coherence (all sets)	600	0 incoherent

True refusal rate: 0 / 400 harmful prompts. De-risked, coherent, and fast — at the same time.

Intended use & disclaimer

This is an uncensored, refusal-ablated model intended for security research, red-teaming, and evaluation by professionals operating under proper authorization. It complies with harmful requests by design — that is what makes it useful as a controlled test artifact, and dangerous if misused. You are responsible for how you use it and for complying with all applicable laws and the base model's license. Provided as-is, without warranty, for legitimate research.

Licensing & access

This is a commercial artifact from Blackfrost-AI. The tuned serving stack and deployment kit (the recipe to reproduce this performance on Blackwell) are provided to licensees, not published here. For commercial licensing, access, or deployment support, contact Blackfrost-AI.

Base model: GLM-5.2 (ZhipuAI / Z.ai), used under the GLM-5.2 license.

Downloads last month: 2

Safetensors

Model size

305B params

Tensor type

BF16

F8_E4M3

Model tree for Blackfrost-AI/GLM-5.2-ABLITERATED-REAP-NU176-NVFP4

Base model

zai-org/GLM-5.2

Quantized

(108)

this model

Collection including Blackfrost-AI/GLM-5.2-ABLITERATED-REAP-NU176-NVFP4

GLM 5.2

Collection

versions of glm 5.2 made for security research 0 refusals. It will ask for assurance but never refuses • 3 items • Updated 4 days ago