Qwen-AgentWorld, a "world model" that simulates agent environments

Alibaba's Qwen-AgentWorld is a 35B MoE world model that simulates seven text and GUI agent environments to train and evaluate AI agents.

Qwen, Alibaba's lab, is launching Qwen-AgentWorld, which it describes as a language-based « world model ». The idea flips the script on agents: models have been extensively trained to act within environments, never to model these environments themselves, that is, to predict what happens next given a state and an agent's action. This is what this model aims for, simulating seven types of environments within a single one: four in text mode (a terminal, a search engine, API servers via MCP, a code environment) and three in graphical user interface (GUI) mode (a browser, an operating system, Android). Environment modeling is the training objective from the outset, via a three-stage pipeline (continuous pre-training, supervised fine-tuning, RL), and not a late graft onto a general-purpose LLM.

Qwen explores two uses. As a decoupled simulator, the model replaces the real environment during an agent's reinforcement training: it predicts the environment's response to each action, which, according to the lab, opens up a scale and control that reality does not allow, even surpassing training conducted in real-world conditions. As a foundation model, a warm-up in state prediction then transfers to multi-turn agent tasks, including outside its training domain.

The open model, an MoE architecture with 35 billion parameters, 3 of which are activated, and a 256K token window, is available on Hugging Face and ModelScope, accompanied by AgentWorldBench, a benchmark across the seven domains based on real-world observations.