Instructions to use ReconAI/Qwen3.5-0.8B-Detection with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use ReconAI/Qwen3.5-0.8B-Detection with Transformers:

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

pipe = pipeline("image-text-to-text", model="ReconAI/Qwen3.5-0.8B-Detection")
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("ReconAI/Qwen3.5-0.8B-Detection")
model = AutoModelForMultimodalLM.from_pretrained("ReconAI/Qwen3.5-0.8B-Detection", device_map="auto")
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]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use ReconAI/Qwen3.5-0.8B-Detection with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "ReconAI/Qwen3.5-0.8B-Detection"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "ReconAI/Qwen3.5-0.8B-Detection",
		"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/ReconAI/Qwen3.5-0.8B-Detection

SGLang

How to use ReconAI/Qwen3.5-0.8B-Detection 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 "ReconAI/Qwen3.5-0.8B-Detection" \
    --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": "ReconAI/Qwen3.5-0.8B-Detection",
		"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 "ReconAI/Qwen3.5-0.8B-Detection" \
        --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": "ReconAI/Qwen3.5-0.8B-Detection",
		"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 ReconAI/Qwen3.5-0.8B-Detection with Docker Model Runner:
```
docker model run hf.co/ReconAI/Qwen3.5-0.8B-Detection
```

Qwen3.5-0.8B-Detection

English | 中文

Qwen3.5-0.8B optimized for object detection. Trained on COCO 2017: detection SFT first, then GSPO reinforcement learning post-training.

Usage is prompt-based: you list the categories you want in the prompt, and the model returns bounding boxes for them as JSON. Categories that are not present in the image are not reported — during RL, categories absent from the image were deliberately mixed into the prompt to train distractor robustness.

Quick start

import torch
from PIL import Image
from transformers import AutoModelForImageTextToText, AutoProcessor

MODEL = "ReconAI/Qwen3.5-0.8B-Detection"

processor = AutoProcessor.from_pretrained(MODEL)
model = AutoModelForImageTextToText.from_pretrained(MODEL, torch_dtype=torch.bfloat16, device_map="auto")

image = Image.open("demo.jpg").convert("RGB")
categories = ["person", "accessory/umbrella", "accessory/handbag"]

prompt = (
    "\n Locate every instance that belongs to the following categories: \n"
    + "".join(f"\t{c}\n" for c in categories)
    + " \nReport bbox coordinates in JSON format."
)
messages = [{"role": "user", "content": [{"type": "image"}, {"type": "text", "text": prompt}]}]
text = processor.apply_chat_template(
    messages, add_generation_prompt=True, tokenize=False, enable_thinking=False
)
inputs = processor(images=[image], text=[text], add_special_tokens=False, return_tensors="pt").to(model.device)

with torch.inference_mode():
    generated = model.generate(**inputs, max_new_tokens=1024, do_sample=False)

response = processor.tokenizer.decode(generated[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True)
print(response)

Output:

```json
[
    {"bbox_2d": [332, 88, 697, 227], "label": "accessory/umbrella"},
    {"bbox_2d": [231, 227, 482, 989], "label": "person"},
    {"bbox_2d": [376, 202, 602, 989], "label": "person"},
    {"bbox_2d": [506, 234, 762, 989], "label": "person"},
    {"bbox_2d": [190, 366, 366, 697], "label": "accessory/handbag"},
    {"bbox_2d": [666, 380, 839, 676], "label": "accessory/handbag"}
]
```

Coordinates

bbox_2d is [x0, y0, x1, y1], normalized to 0–1000, with x normalized by image width and y by image height independently. To convert back to pixels:

x_px = x / 1000 * image.width
y_px = y / 1000 * image.height

Training

Stage	Setup
SFT	Full COCO 2017 train detection annotations
RL	GSPO, 600 steps, reward = soft-count F1 (IoU as partial credit)

On the validation set (first 200 images of COCO val2017), soft-F1 improved from 0.437 to 0.479 over RL. The gain comes mainly from precision: the average number of predicted boxes dropped from 10.5 to 6.2 while the ground-truth count stayed around 6.5 — the model learned to stop over-reporting.

sft training on coco2017

gspo training on coco2017

Limitations

Vocabulary is limited to the COCO 80 categories; names outside it were never trained on
A 0.8B autoregressive VLM — weaker than dedicated detectors (DETR / YOLO family) on dense small objects and heavy occlusion
Single-image input, output capped at 1024 tokens