Instructions to use buugiiiiii/GLM-4.7-Flash-KGuard-Soft-56E-Control with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use buugiiiiii/GLM-4.7-Flash-KGuard-Soft-56E-Control with Transformers:

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

pipe = pipeline("text-generation", model="buugiiiiii/GLM-4.7-Flash-KGuard-Soft-56E-Control")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("buugiiiiii/GLM-4.7-Flash-KGuard-Soft-56E-Control")
model = AutoModelForCausalLM.from_pretrained("buugiiiiii/GLM-4.7-Flash-KGuard-Soft-56E-Control", 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=256)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use buugiiiiii/GLM-4.7-Flash-KGuard-Soft-56E-Control with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "buugiiiiii/GLM-4.7-Flash-KGuard-Soft-56E-Control"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "buugiiiiii/GLM-4.7-Flash-KGuard-Soft-56E-Control",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/buugiiiiii/GLM-4.7-Flash-KGuard-Soft-56E-Control

SGLang

How to use buugiiiiii/GLM-4.7-Flash-KGuard-Soft-56E-Control 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 "buugiiiiii/GLM-4.7-Flash-KGuard-Soft-56E-Control" \
    --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": "buugiiiiii/GLM-4.7-Flash-KGuard-Soft-56E-Control",
		"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 "buugiiiiii/GLM-4.7-Flash-KGuard-Soft-56E-Control" \
        --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": "buugiiiiii/GLM-4.7-Flash-KGuard-Soft-56E-Control",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use buugiiiiii/GLM-4.7-Flash-KGuard-Soft-56E-Control with Docker Model Runner:
```
docker model run hf.co/buugiiiiii/GLM-4.7-Flash-KGuard-Soft-56E-Control
```

GLM-4.7-Flash K-Guard Soft 56E Control

Public research-control checkpoint from the GLM-4.7-Flash K-Guard REAP study.

Routed experts: 64 → 56 per MoE layer (12.5% pruning)
Active experts per token: top-4 unchanged
Base revision: 7dd20894a642a0aa287e9827cb1a1f7f91386b67
Role: soft-protection control
Fine-tuning: none
Precision: BF16
Status: experimental; release_eligible=false

This checkpoint is published for reproducibility and controlled comparison. It is not the strict K-Guard 56E release candidate. It was evaluated in B22 and later selected over a new KO85 candidate on the separate B25 development set; because B25 retained the existing model, B26 correctly did not reinterpret its earlier held-out results as a new held-out run.

B22 controlled benchmark

All models below use the same 56E budget. Values are percentages. LiveCodeBench uses v6, greedy n=1, so it is not directly interchangeable with official n=10 leaderboard scores.

Metric	Base 64E	SameCal REAP 56E	Soft K-Guard 56E
Korean MC macro	68.79	66.79	66.79
Academic 10-benchmark macro	65.51	64.51	64.97
Coding 3-benchmark macro	58.28	57.54	58.38

The Korean macro difference between Soft K-Guard and SameCal REAP was approximately 0.00 pp with a paired stratified bootstrap 95% interval of [-1.02, 1.03]. This run therefore does not establish a Korean causal advantage from soft protection. It is a useful control, not evidence of a general improvement.

Loading

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "buugiiiiii/GLM-4.7-Flash-KGuard-Soft-56E-Control"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)

Use a recent Transformers version that supports Glm4MoeLiteForCausalLM.

Limitations

One calibration run and one 56E budget; calibration-seed stability is unknown.
The checkpoint is not production-certified and has no fine-tuning recovery.
Top-4 routing is unchanged, so checkpoint reduction does not by itself prove higher decode throughput.
The public weights support reproducibility; the experiment code and detailed manifests remain in the associated private research repository.

Downloads last month: 14

Safetensors

Model size

28B params

Tensor type

BF16

F32

Model tree for buugiiiiii/GLM-4.7-Flash-KGuard-Soft-56E-Control

Base model

zai-org/GLM-4.7-Flash

Finetuned

(86)

this model