Update README.md

README.md

@@ -1,3 +1,118 @@
----
-license: apache-2.0
----
+---
+license: apache-2.0
+language:
+- en
+base_model:
+- Qwen/Qwen3-VL-32B-Thinking
+library_name: transformers
+tags:
+- vlm
+- web-agent
+- opagent
+---
+# OpAgent-32B
+
+<!-- Provide a quick summary of what the model is/does. -->
+
+**OpAgent-32B** is a powerful, open-source Vision-Language Model (VLM) specifically fine-tuned for autonomous web navigation. It serves as the core single-model engine within the broader **[OpAgent project](https://github.com/codefuse-ai/OpAgent)**.
+
+
+
+## Model Details
+
+### Model Description
+
+<!-- Provide a longer summary of what this model is. -->
+
+- **Base Model:** `Qwen3-VL-32B-Thinking`
+- **Fine-tuning Strategy:** Hierarchical Multi-Task SFT followed by Online Agentic RL with a Hybrid Reward mechanism.
+- **Primary Task:** Autonomous web navigation and task execution.
+- **Input:** A combination of a natural language task description and a webpage screenshot.
+- **Output:** A JSON-formatted action (e.g., `click`, `type`, `scroll`) or a final answer.
+
+### Model Sources [optional]
+
+<!-- Provide the basic links for the model. -->
+
+- **Repository:** [https://github.com/codefuse-ai/OpAgent]
+
+
+
+## Uses
+
+This model is designed to be used as a web agent. The primary way to run it is through a high-performance inference engine like **vLLM**, as demonstrated in our [single-model usage guide](https://github.com/codefuse-ai/OpAgent/tree/main/opagent_single_model).
+
+Below is a Python code snippet demonstrating how to use `OpAgent-32B` with `vLLM` for a single-step inference.
+
+```python
+import base64
+from vllm import LLM, SamplingParams
+from PIL import Image
+from io import BytesIO
+
+# --- 1. Helper function to encode image ---
+def encode_image_to_base64(image_path):
+    with Image.open(image_path) as img:
+        buffered = BytesIO()
+        img.save(buffered, format="PNG")
+        return base64.b64encode(buffered.getvalue()).decode('utf-8')
+
+# --- 2. Initialize the vLLM engine ---
+# Ensure you have enough GPU memory.
+model_id = "codefuse-ai/OpAgent-32B"
+llm = LLM(
+    model=model_id,
+    trust_remote_code=True,
+    tensor_parallel_size=1,  # Adjust based on your GPU setup
+    gpu_memory_utilization=0.9
+)
+
+# --- 3. Prepare the prompt ---
+# The prompt must include the system message, task description, and the screenshot.
+task_description = "Search for wireless headphones under $50"
+screenshot_path = "path/to/your/screenshot.png" # Replace with your screenshot path
+base64_image = encode_image_to_base64(screenshot_path)
+
+# This prompt format is crucial for the agent's performance
+prompt = f"""system
+You are a helpful web agent. Your goal is to perform tasks on a web page based on a screenshot and a user's instruction.
+Output the thinking process in <think> </think> tags, and for each function call, return a json object with function name and arguments within <tool_call></tool_call> XML tags as follows:\n<think> ... </think><tool_call>{"name": <function-name>, "arguments": <args-json-object>}</tool_call>.
+user
+[SCREENSHOT]
+Task: {task_description}
+assistant
+"""
+
+# --- 4. Generate the action ---
+sampling_params = SamplingParams(temperature=0.0, max_tokens=1024)
+
+# The model expects the image to be passed via the `images` parameter
+outputs = llm.generate(
+    prompts=[prompt],
+    sampling_params=sampling_params,
+    images=[base64_image]
+)
+
+# --- 5. Print the result ---
+for output in outputs:
+    generated_text = output.outputs[0].text
+    print("--- Generated Action ---")
+    print(generated_text)
+
+```
+
+For a complete, interactive agent implementation, please see the code in the [`opagent_single_model`](https://github.com/codefuse-ai/OpAgent/tree/main/opagent_single_model) directory of our repository.
+
+## Citation
+
+If you use `OpAgent-32B` or the `OAgent` framework in your research, please cite our work:
+
+```bibtex
+@misc{opagent2026,
+  author = {CodeFuse-AI Team},
+  title = {OpAgent: Operator Agent for Web Navigation},
+  year = {2026},
+  publisher = {GitHub},
+  howpublished = {\url{https://github.com/codefuse-ai/OpAgent}},
+  url = {https://github.com/codefuse-ai/OpAgent}
+}