Qwen-AgentWorld Language World Model

Skill by ara.so — AI Agent Skills collection.

What It Does

Qwen-AgentWorld is a native language world model that simulates agentic environments across seven unified domains: MCP (tool-calling), Search, Terminal, SWE (software engineering), Android, Web, and OS. Unlike traditional approaches that adapt language models post-hoc, Qwen-AgentWorld is trained from Continued Pre-Training (CPT) onward with environment modeling as the core objective.

Key Capabilities:

• Unified multi-domain simulation: Single model covers 7 agent interaction environments
• Controllable simulation: Inject perturbations, create fictional worlds, adapt environments
• Agent foundation model: Sim RL warm-up transfers to multi-turn, tool-calling agentic tasks
• Zero-shot generalization: Out-of-distribution environment handling (e.g., Claw Agent)
• Evaluation benchmark: AgentWorldBench with 5-dimensional rubric scoring

Model Variants:

• Qwen-AgentWorld-35B-A3B (35B total, 3B active MoE, 256K context)
• Qwen-AgentWorld-397B-A17B (397B total, 17B active MoE, 256K context)