Instructions to use Qwen/Qwen3.6-27B with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use Qwen/Qwen3.6-27B with Transformers:

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

pipe = pipeline("image-text-to-text", model="Qwen/Qwen3.6-27B")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
pipe(text=messages)

# Load model directly
from transformers import AutoProcessor, AutoModelForMultimodalLM

processor = AutoProcessor.from_pretrained("Qwen/Qwen3.6-27B")
model = AutoModelForMultimodalLM.from_pretrained("Qwen/Qwen3.6-27B")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
inputs = processor.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(processor.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Inference
HuggingChat
Notebooks
Google Colab
Kaggle
AMD Developer Cloud
Local Apps Settings

vLLM

How to use Qwen/Qwen3.6-27B with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "Qwen/Qwen3.6-27B"
# 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/Qwen3.6-27B",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Use Docker

docker model run hf.co/Qwen/Qwen3.6-27B

SGLang

How to use Qwen/Qwen3.6-27B 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/Qwen3.6-27B" \
    --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/Qwen3.6-27B",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

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/Qwen3.6-27B" \
        --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/Qwen3.6-27B",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Docker Model Runner
How to use Qwen/Qwen3.6-27B with Docker Model Runner:
```
docker model run hf.co/Qwen/Qwen3.6-27B
```

Dense 60-120B

#44

by rtzurtz - opened 27 days ago

Discussion

rtzurtz

27 days ago

•

edited 21 days ago

Qwen3.6-27B says that to see a significant improvement, a minimum target of 60-80B parameters would be required and an ideal target, for major gains, would be 100-130B.

Motivation:

High RAM/memory prices, especially RAM.
A dense model provides a much better performance for a given parameter size (see¹ for what it's worth) vs a MoE model.
My 70 GB (64 GB RAM + 12 GB VRAM - 6 GB for the OS + x GB for the context) can't fit a good, UD-Q4_K_M, 122B quant, but can easily fit the better performing (at least according to¹ and also various comments) 27B quant.
For me and from comments, for many, speed can be much less relevant for a high quality answer.
That being said, a 60B Q4_K_M quant (approx. 40 GB) will still fit in a 2 x 24 GB setup and run pretty fast.
The 6.4 times (5984744/932498) higher monthly download count for the 27B dense vs 122B MoE may confirm the limited memory for many (yes, the 27B is the newer Qwen3.6, but the 3.5-122B has never been as popular as 3.5-27B (2452527 monthly downloads)).

¹ https://artificialanalysis.ai/?models=qwen3-6-27b-non-reasoning%2Cqwen3-5-122b-a10b%2Cqwen3-5-122b-a10b-non-reasoning%2Cqwen3-6-27b%2Cqwen3-6-35b-a3b-non-reasoning%2Cqwen3-6-35b-a3b&intelligence=artificial-analysis-intelligence-index

wojciechslusarczyk

3 days ago

•

edited 3 days ago

Well… the thing is that a 27B dense model will reliably outperform a 35B MoE (or larger) model as long as the payload stays within a reasonable range — say, up to ~128k tokens. Once you cross that line, a small dense model may (or may not) start losing a bit of attention stability (minor language slips, small factual inconsistencies, etc.) Not because it’s “stupid”, but simply because it’s too small to maintain a perfectly stable attention matrix at very large context sizes — mathematics always wins.

Moreover, in real‑world use I would never choose a Q4‑quantized 60B over a smaller dense BF16/FP16 model for strictly logical, multi‑layer reasoning tasks where a high-quality answer is expected. Quantization noise (plus MoE routing noise) accumulates, and deep causal reasoning suffers.

A genuinely intelligent model (see Qwen3.6 27B) doesn’t need a full Wikipedia baked into its weights — it can compensate for missing knowledge through external tools. Intelligence is the most valuable property of Qwen3.6, unless the model is used purely for casual chatbot tasks.

Regards,
/WS

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