Alibaba's Qwen launches Qwen-AgentWorld, the first native language world model

Alibaba's Qwen launched Qwen-AgentWorld today, calling it the first native language world model. This model treats environment modeling as a training goal from the start, going through the full process of CPT, SFT, and RL, rather than adapting a general-purpose large language model later.

Qwen-AgentWorld is a single model that covers both text-based environments (like MCP, Search, Terminal, SWE) and GUI-based environments (like Web, OS, Android), allowing knowledge to transfer across different areas. It comes in two sizes: 35B-A3B and 397B-A17B, trained on over 10 million real-world interaction records.

Alongside it, the AgentWorldBench evaluation benchmark covers seven areas, with each test sample including real-world execution data. In tests, Qwen-AgentWorld-397B-A17B achieved the highest overall simulation quality, outperforming GPT-5.4, Claude Opus 4.8, and Gemini 3.1 Pro. The model and benchmark are available on Hugging Face and ModelScope.