Qwen's AgentWorld-35B-A3B simulates 7 agent environments with 3B active params

Qwen just released AgentWorld-35B-A3B, a 35B-parameter mixture-of-experts (MoE) language world model that activates only about 3B parameters per token. Unlike standard chat models or autonomous agents, this model is explicitly designed to simulate the environment side of an agent loop: given a history of actions and a new tool or GUI action, it predicts the next observation or state. This makes it a unique asset for agent development and evaluation.

The model covers seven distinct agent interaction domains: MCP (Model Context Protocol) and tool calling, search, terminal commands, software engineering (SWE), Android, web browsing, and OS-level GUI interactions. By predicting environment feedback rather than requiring real tool execution, AgentWorld can generate synthetic trajectories for training, enable offline evaluation without latency or cost, and serve as a sandbox for testing tool-use workflows.

This approach addresses a key bottleneck in agent development: the expense and complexity of running real environments at scale. With AgentWorld, developers can simulate thousands of agent interactions cheaply, iterate on prompts and policies faster, and benchmark performance across diverse domains. The model's MoE architecture keeps inference efficient despite its large total parameter count, making it practical for research and production pipelines.

Qwen's release signals a growing focus on environment simulation as a critical infrastructure for agentic AI. By providing a single model that spans so many interaction modalities, AgentWorld-35B-A3B could accelerate progress in areas like autonomous software development, robotic control, and general-purpose digital assistants. The Hugging Face model is open for community experimentation.