---
title: AgentWorld Demo
emoji: 🌍
colorFrom: blue
colorTo: green
sdk: gradio
sdk_version: 6.0.0
app_file: app.py
pinned: false
license: apache-2.0
---

# 🌍 AgentWorld Demo — Language World Model Simulator

A Gradio webapp demonstrating the capabilities of **Qwen-AgentWorld**, a Language World Model that simulates agentic environments. Instead of running real terminals, browsers, or APIs, the model *predicts* what those environments would return — enabling fast, cheap, and safe agent testing.

## What it simulates

| Domain | Description |
|---|---|
| 🖥 **Terminal** | Linux shell — predicts command output, tracks filesystem state across turns |
| 🔍 **Search** | Web search engine — simulates search results, page extraction, memory store |
| 🔧 **MCP / Tools** | Tool-calling APIs — simulates CRUD operations, API responses, error handling |
| 🌐 **Web** | Browser — predicts page state after clicks, navigation, form fills |

## How it works

1. Pick a domain from the dropdown
2. Enter an action in JSON format (click "Load Example" to get started)
3. The model predicts what the environment would return
4. Multi-turn: the model tracks state across your session (files created, pages visited, etc.)

## Model

This Space uses **Qwen2.5-7B-Instruct** (free HF inference) with the official AgentWorld system prompts from the [Qwen-AgentWorld repository](https://github.com/QwenLM/Qwen-AgentWorld).

For full simulation fidelity, deploy the actual [Qwen-AgentWorld-35B-A3B](https://huggingface.co/Qwen/Qwen-AgentWorld-35B-A3B) model on a paid Inference Endpoint and set `INFERENCE_ENDPOINT_URL`.

## Run locally

```bash
pip install -r requirements.txt
python app.py
```

## Credits

- [Qwen-AgentWorld](https://huggingface.co/Qwen/Qwen-AgentWorld-35B-A3B) by the Qwen team
- System prompts from [QwenLM/Qwen-AgentWorld](https://github.com/QwenLM/Qwen-AgentWorld)
- Built by [geraldamasi](https://huggingface.co/geraldamasi)