Qwen releases an open agent world model

Alibaba’s Qwen team released Qwen-AgentWorld on June 24, a model that predicts what an environment returns after an agent acts. Instead of choosing actions, it simulates the response: what the terminal prints, what the browser loads, what the Android screen shows. One model covers seven domains, MCP, search, terminal, software engineering, web, operating system, and Android, under a single training objective.

The 35-billion-parameter variant and the AgentWorldBench test set are out under the Apache 2.0 license. The flagship 397-billion-parameter version, which uses 17 billion active parameters, posts the highest simulation quality on AgentWorldBench, ahead of GPT-5.4 and Claude Opus 4.8 in the team’s comparison.

What it means for you

Testing an agent against live systems is slow and risky: real API calls, real side effects, real bills. A world model lets you rehearse an agent in a simulated copy of its environment first, so you catch failures before they touch anything that matters.

It is also one more open release aimed squarely at builders, the same direction as Ornith’s open coding models and the broader case that open models are now good enough for real operator work. If agents are new to you, start with coding agents for non-developers.