Tongyi Qianwen Launches Native Language World Model Qwen-AgentWorld
Published · Jun 24 · Wed Source · 千问大模型 (CN)

Tongyi Qianwen Launches Native Language World Model Qwen-AgentWorld

Alibaba Tongyi officially launches Qwen-AgentWorld, the first native language world model, capable of simulating agentic interaction environments across seven major domains: MCP, Search, Terminal, SWE, Web, OS, and Android. The model is based on over 10 million real-world environment trajectories and undergoes three-stage end-to-end training via CPT→SFT→RL. On the self-developed evaluation benchmark AgentWorldBench, its overall simulation quality surpasses top models such as GPT-5.4 and Claude Opus 4.8.

KeywordsGPTClaudeQwenAgentTongyiQianwenLaunchesNative

Alibaba Tongyi officially launches Qwen-AgentWorld, the first native language world model, capable of simulating agentic interaction environments across seven major domains: MCP, Search, Terminal, SWE, Web, OS, and Android. The model is based on over 10 million real-world environment trajectories and undergoes three-stage end-to-end training via CPT→SFT→RL. On the self-developed evaluation benchmark AgentWorldBench, its overall simulation quality surpasses top models such as GPT-5.4 and Claude Opus 4.8.

June 25, 2026 • Qwen-AgentWorld is a native language world model that simulates agentic environments via long chain-of-thought reasoning across seven unified domains: MCP, Search, …

June 24, 2026 • Full leaderboard (Overall descending) ... Core conclusion: 🏆 Qwen-AgentWorld-397B-A17B overall score 58.71, surpassing GPT-5.4 (58.25), topping the list 🚀 Qwen-AgentWorld-35B-A3B at the same 35B scale, compared to base …

June 24, 2026 • Figure 1: Qwen-AgentWorld overview diagram, showing its world model training process covering seven major domains, as well as its dual paradigm as a new type of environment simulator and agentic foundation. From the diagram, we can see that the training data scale of this model …

June 23, 2026 • First, as a decoupled environment simulator, Qwen-AgentWorld supports scalable and controllable simulation of thousands of real-world environments for agentic RL, yielding gains that ….

This page provides an editorial summary based on publicly available information. It is not a republished article. Use the source link below for the original report.