Qwen-AgentWorld: Language World Models for General Agents

TL;DR

Qwen just released Qwen-AgentWorld, a language model trained to simulate agentic environments. Instead of training agents in real environments (slow, expensive, uncontrollable), train a model that can predict what the environment would return for any given action. The model is the environment.

It covers 7 domains in a single model: MCP, Search, Terminal, Android, Web Browser, OS, and SWE.

Model Sizes

• Qwen-AgentWorld-35B-A3B — MoE, 35B total / 3B active, 256K context — open source
• Qwen-AgentWorld-397B-A17B — not open source

Three-Stage Training Pipeline

The paper "Language World Models for General Agents" describes a three-stage process:

1. CPT (Continued Pre-Training) — 10M+ real-world interaction trajectories across 7 domains. The model learns state transition dynamics.
2. SFT (Supervised Fine-Tuning) — Activates next-state-prediction reasoning via long chain-of-thought.
3. RL (Reinforcement Learning) — Hybrid rubric-and-rule reward to sharpen simulation fidelity.

The key difference from prior work: environment modeling is the core training objective from stage one, not a post-hoc add-on. Hence "native world model."

Benchmark Results

They created AgentWorldBench, spanning 7 domains with data from 5 frontier models.

The 35B-A3B version (3B active parameters) matches Claude Opus 4.8 within 1 point. Without world model training, the same architecture scores only 47.73 — LWM training provides a +8.66 boost.

Key Discovery 1: Simulated Training Beats Real Training

Using Qwen-AgentWorld as a simulated environment for RL actually outperforms training in real environments.

It also supports controllable simulation: inject perturbations, construct fictional worlds, train under harder conditions. Agents trained in fictional search worlds still generalize to real search (F1 +16.29).

Key Discovery 2: Predicting Environments Makes Agents Stronger

LWM warm-up on single-turn, non-agentic trajectories improves downstream agent performance — even on out-of-domain tasks with zero agent-specific fine-tuning.

Why This Matters

This is the AlphaGo approach applied to agents: self-play in a simulated world, without needing real environments. If it scales, agent training costs drop dramatically — one model replaces thousands of browser instances, API calls, and Docker containers.

The open question: how well does the simulation generalize? The gap between simulated and real environments still exists. But as a proof of concept, this is one of the most interesting agent research directions of 2026.

Resources

1. Paper: https://arxiv.org/abs/2606.24597
2. GitHub: https://github.com/QwenLM/Qwen-AgentWorld
3. Model: https://huggingface.co/Qwen/Qwen-AgentWorld-35B-A3B
4. Benchmark: https://huggingface.co/datasets/Qwen/AgentWorldBench
5. X announcement: https://x.com/Alibaba_Qwen/status/2069720365442719867