Tongyi Qianwen has released its first native language world model, Qwen-AgentWorld.

According to monitoring by Dongcha Beating, the Qwen team today officially launched Qwen-AgentWorld — the first native language world model that sets environment modeling as a training objective from the continued pre-training phase, rather than an after-the-fact adaptation of a general large language model. The model covers seven domains under a unified framework: text-based environments (MCP, Search, Terminal, SWE) and GUI-based environments (Web, OS, Android). Trained on over 10 million real-world interaction trajectories via three stages — CPT → SFT → RL — it enables cross-domain knowledge transfer. Also open-sourced simultaneously is the AgentWorldBench evaluation benchmark, with each test sample paired with observation data collected from real environment executions. Both the model and the benchmark are now available on Hugging Face and ModelScope. In the AgentWorldBench evaluation, Qwen-AgentWorld-397B-A17B achieved the highest overall simulation quality, outperforming GPT-5.4, Claude Opus 4.8, and Gemini 3.1 Pro. The research team also explored two application paths for world modeling in agent training: First, as a decoupled environment simulator, controllable simulated reinforcement learning (RL) can shape agent behavior and significantly outperforms RL trained exclusively in real environments. Second, as a unified agent base model, LWM warm-up training can be transferred to multi-turn agent tasks across seven benchmarks, three of which are not included in the training set at all, and no RL fine-tuning for agent tasks is required. The work preliminarily validates the potential of language world models as stronger foundational models for agents. Click the original link below to join the Dongcha Beating · Feishu AI News channel, which monitors global AI hotspots and news around the clock.