Qwen today released Qwen-AgentWorld, billed as the first native Language World
Model (LWM) that can simulate agent interaction environments across seven
domains. Native world modeling: environment modeling is an explicit training
objective beginning in continued pretraining (CPT) and maintained through CPT →
supervised fine-tuning (SFT) → reinforcement learning (RL), rather than as a
post‑hoc adaptation of a general LLM. Seven domains, one model: a single model
covers text environments (MCP, Search, Terminal, SWE) and GUI environments (Web,
OS, Android), enabling cross‑domain knowledge transfer.
