Qwen/Qwen-AgentWorld-35B-A3B

Qwen-AgentWorld-35B-A3B: A Native Language World Model

Qwen-AgentWorld-35B-A3B is a 35 billion parameter language world model (LWM) developed by Qwen, specifically engineered for simulating agentic environments. Unlike general-purpose LLMs adapted for world modeling, this model is a native world model, with environment modeling as its core training objective from the initial Continual Pre-Training (CPT) stage. It covers seven distinct agent interaction domains: MCP (tool calling), Search, Terminal, SWE (software engineering), Android, Web, and OS, supporting both text and GUI interactions.

Key Capabilities

• Unified Domain Coverage: Simulates environments across seven diverse domains within a single model.
• Long Context Understanding: Features an extensive 262,144 token context length, crucial for multi-turn environment simulation.
• Generalizable Simulator: Demonstrates zero-shot generalization to out-of-domain environments, such as OpenClaw.
• Agent Foundation Model: Its LWM RL warm-up transfers effectively to multi-turn, tool-calling agentic tasks across multiple benchmarks.
• Three-Stage Training: Utilizes a CPT, Supervised Fine-Tuning (SFT), and Reinforcement Learning (RL) pipeline to enhance simulation fidelity.

Good For

• Agent Development and Testing: Ideal for simulating complex agentic environments to develop and evaluate AI agents.
• Environment Simulation: Excels at predicting the next environment state given an agent's action and interaction history.
• Research in Language World Models: Provides a robust platform for exploring and advancing LWM capabilities, especially in agentic settings.