Instructions to use Nanbeige/Nanbeige4.2-3B with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use Nanbeige/Nanbeige4.2-3B with Transformers:

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

pipe = pipeline("text-generation", model="Nanbeige/Nanbeige4.2-3B", trust_remote_code=True)
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained("Nanbeige/Nanbeige4.2-3B", trust_remote_code=True, device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use Nanbeige/Nanbeige4.2-3B with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "Nanbeige/Nanbeige4.2-3B"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Nanbeige/Nanbeige4.2-3B",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/Nanbeige/Nanbeige4.2-3B

SGLang

How to use Nanbeige/Nanbeige4.2-3B 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 "Nanbeige/Nanbeige4.2-3B" \
    --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": "Nanbeige/Nanbeige4.2-3B",
		"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 "Nanbeige/Nanbeige4.2-3B" \
        --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": "Nanbeige/Nanbeige4.2-3B",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use Nanbeige/Nanbeige4.2-3B with Docker Model Runner:
```
docker model run hf.co/Nanbeige/Nanbeige4.2-3B
```

Q on the model

by TomLucidor - opened 3 days ago

Discussion

TomLucidor

3 days ago

Can you 4-bit quantize this model but still endure its quality isn't bad
Will there be bigger models using the looped architecture
Would you make MoE and/or linear attention and/or SWA models next
What is your stance on LFM and other newer architectures
Will there be DFlash or Eagele3 or MTP for this model

leran1995

Nanbeige LLM Lab org 3 days ago

The main purpose of the looped architecture is to improve model capacity without increasing the parameter count. We believe this tradeoff is particularly well suited to the deployment and application scenarios of smaller models.

We are also actively exploring MoE and linear/sparse attention architectures. Related releases are planned not long after the ongoing Nanbeige4.5. 😊

TomLucidor

3 days ago

@leran1995 what are the major acceleration methods on the roadmap that makes both parameter efficiency and compute efficiency (and interpretability with J-Lens) possible?

owao

2 days ago

@TomLucidor just in case, for dflash or eagle, you don't need the model to already have a native feature (as you need for mtp). vllm have full tutos and notebooks for each of the methods, if you are interested https://docs.vllm.ai/projects/speculators/en/latest/user_guide/tutorials/ and have enough ram ;) personnaly I don't even have enough ram to COMPILE vllm (the Nanbeige custom one) xDD

TomLucidor

1 day ago

@leran1995 just to write more on the matter, is it even possible to distill regular models like Qwen3.5 and Gemma4 directly into Nanbeige (compaction and compute-minimization)?

Upload images, audio, and videos by dragging in the text input, pasting, or clicking here.

Tap or paste here to upload images

· Sign up or log in to comment