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, AutoModelForImageTextToText

processor = AutoProcessor.from_pretrained("Qwen/Qwen3.6-27B")
model = AutoModelForImageTextToText.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
```

90.0% on SWE-bench Verified using Qwen3.6-27B-FP8 — open-weight, consumer GPUs, full transparency

#33

by mrguo - opened 26 days ago

Discussion

mrguo

26 days ago

Hi Qwen team and HF community 👋

Wanted to share a result that's been three months in the making.

TL;DR: Qwen3.6-27B-FP8 (your published weights, sglang-served with served-name Qwen3.5-27B-Thinking for legacy compatibility — same checkpoint) reaches 90.0% on SWE-bench Verified — competitive with closed-API SOTA — running on 12× consumer RTX 4090s modded to 48GB. No H100 / A100.

Setup

Model: Qwen3.6-27B-FP8, Alibaba's official FP8 release, unmodified (no fine-tune, no distillation, no continual training)
Inference: sglang, 3 tensor-parallel instances of 4 GPUs each
Agent stack: @anthropic-ai/claude-code@2.1.23 in -p one-shot mode, behind a custom protocol-translation proxy (~47K LoC Python)
Hardware: 12 × RTX 4090 modded 48GB across 2 workstations. Minimum reproducible config: 1 workstation × 4 cards.

Results (two-tier honest disclosure)

90.0% headline (450/500) — single-attempt + 13-instance cli/proxy crash retry (6 recovered) + 4 psf__requests instances re-evaluated against real httpbin.org
88.0% strict floor (440/500) — zero retry, zero env fix

Both preds files + official swebench.harness reports + SHA256SUMS are public.

Anti-cheat ablation (the part I think the Qwen team will find most interesting)

Same model + same proxy + Alibaba's own qwen-code CLI instead of claude-cli → 87.4% headline / 85.4% strict floor. Both significantly clear public SOTA (79.2%).

The ~2.6 pt gap between CLIs is consistent across all tiers — structural (tool-surface design), not dataset-specific exploitation. The engineering uplift is CLI-independent.

Baseline comparison

Same Qwen3.6-27B on mini-swe-agent: 67.8%
Same model on engineered agent stack: 90.0%

+22.2 pt from agent engineering alone, with the model held constant.

Why posting here

Public thank-you to the Qwen team — the open-weight 27B FP8 release made this entire experiment possible on a budget an independent researcher can actually afford. The model is the foundation; engineering is leverage.
Reviewer invitation — full repo is public (preds, reports, SHA256SUMS, CLI-swap ablation evidence). Issues open for technical scrutiny.
A data point that the open-weight ceiling is higher than public benchmarks suggest — most public agent benchmarks use simple stacks that underestimate the underlying model.

Honest caveat

The protocol-translation proxy is closed-source. Capability disclosure is in proxy_capabilities.md; source review available to SWE-bench official reviewers under Review Terms. The CLI-swap ablation directly addresses the "task-specific exploitation" concern.

Qwen
/

Qwen3.6-27B