Instructions to use Qwen/Qwen3-8B with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use Qwen/Qwen3-8B with Transformers:

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

pipe = pipeline("text-generation", model="Qwen/Qwen3-8B")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3-8B")
model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-8B")
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]:]))

Inference
HuggingChat
Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use Qwen/Qwen3-8B with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "Qwen/Qwen3-8B"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Qwen/Qwen3-8B",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/Qwen/Qwen3-8B

SGLang

How to use Qwen/Qwen3-8B 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 "Qwen/Qwen3-8B" \
    --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": "Qwen/Qwen3-8B",
		"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 "Qwen/Qwen3-8B" \
        --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": "Qwen/Qwen3-8B",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use Qwen/Qwen3-8B with Docker Model Runner:
```
docker model run hf.co/Qwen/Qwen3-8B
```

DoRA personal-voice adapter on Qwen3-8B — 100% blind A/B vs stock, $1.50 GPU cost (writeup + adapter)

#34

by yukakst - opened 9 days ago

Discussion

yukakst

9 days ago

Hi Qwen team and community,

Sharing a small experiment we ran on Qwen3-8B that might be of interest to folks doing personalization work.

We trained a DoRA adapter on 6128 personal Telegram message pairs (single person's chat history, with consent). Single RTX 3090 on Vast.ai, 3.5h, ~$1.50 total.

Setup

Base: Qwen/Qwen3-8B
Method: DoRA via peft (use_dora=True), rank 16, alpha 32, target q/k/v/o
Loss mask only on assistant tokens (critical for voice work)
bf16 + flash-attn-2 + gradient checkpointing, peaks at 22/24 GB

Eval (blind 3-way A/B on 30 hold-out prompts)

Comparison	Result
DoRA vs stock (head-to-head)	DoRA 100%
Full 3-way (real human / DoRA / stock)	Real 71% / DoRA 29% / Stock 0%
Catastrophic forgetting on 50 baseline tasks	0 pp

Critical gotcha — enable_thinking=False

This bit us hard. Qwen3 emits <think> reasoning traces by default, but chat-style training data has none. During inference the base prior fights the adapter and output ends up as long reasoning prefix + short colloquial reply. The fix:

inputs = tokenizer.apply_chat_template(
    messages,
    add_generation_prompt=True,
    enable_thinking=False,  # MANDATORY for chat-style adapters
    return_tensors='pt',
)

Worth a prominent note in the Qwen3 README/usage docs IMO — this is non-obvious for anyone training chat adapters and silently breaks output quality.

transformers version

Qwen3 lands in 4.51, but 4.55+ wants torch ≥2.5. We pinned to transformers==4.53.0 for Vast 3090 compatibility.

Adapter on HF (CC BY-NC 4.0, gated for ethical reasons since training data is one specific person):
https://huggingface.co/aiconiccompany/yuka-dora-v1

Full write-up with methodology, configs, the one prompt where DoRA beat the real human at sounding like themselves, and what we'd change in v2:
https://aiconic.company/en/journal/dora-personal-voice

Happy to share configs/eval scripts or discuss with anyone planning similar work on Qwen3.

Thanks to the Qwen team for shipping such a clean open-weights model — the personal-AI thesis only works because of releases like this.

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