What Makes a Language World Model Different

Every AI agent operates in a loop: it takes an action, and the environment responds. Until now, language models were trained to generate text, then retrofitted with tool-calling abilities. Qwen-AgentWorld flips the script: environment modeling is the training objective from day one — through continual pre-training, supervised fine-tuning, and reinforcement learning — across more than 10 million real-world interaction trajectories.

A traditional LLM agent learns what to do. A language world model learns what happens next. Given the interaction history and an agent's action, it predicts the terminal output, the API response, the updated DOM — using multi-step causal reasoning, stateful tracking, and deep domain knowledge. This is not template-based generation; it's faithful environment simulation.

Seven Domains, One Model

Qwen-AgentWorld covers an unprecedented range of environments within a single architecture:

Text-based environments:

• Terminal — shell output, file system state, process behavior
• Search — search engine results with realistic URLs, snippets, and rankings
• MCP — API server responses, database state, service protocol consistency
• SWE — IDE environment: git diffs, test results, compilation errors

GUI environments:

• Web — browser DOM and accessibility tree state changes
• Android — UI hierarchy changes after touch and gesture actions
• OS — desktop file system, window management, application behavior

For GUI environments, the model works with renderable code (accessibility tree XML, HTML, UI hierarchy markup) rather than pixel frames — enabling text-only world modeling of visual domains.

The Training Pipeline: CPT → SFT → RL

The three-stage recipe follows a clear principle: CPT injects, SFT activates, RL sharpens.

Stage 1 — Continual Pre-Training (CPT): The model absorbs environment knowledge from dedicated infrastructure (containerized sandboxes, MCP servers, Android/web/OS emulators), open interaction traces, and in-house agent trajectories. Beyond environment data, it incorporates specialized-domain corpora spanning industrial control, cybersecurity, law, medicine, finance, and current affairs. A key innovation is turn-level information-theoretic loss masking — four surface-level statistics per (action, observation) pair identify information-rich turns while masking the rest from the loss function, dramatically improving training efficiency.

Stage 2 — Supervised Fine-Tuning (SFT): Next-state prediction is activated as an explicit reasoning pattern using <think> blocks. Rejection sampling selects 7,094 high-quality thinking trajectories from the CPT model.

Stage 3 — Reinforcement Learning (RL): Hybrid rewards combine a rubric-based LLM judge (evaluating format, factuality, consistency, realism, and quality) with rule-based verifiers for domains where exact correctness can be checked programmatically — using GSPO for training.

AgentWorldBench: A New Standard for World Models

Alongside the model, the team released AgentWorldBench — a comprehensive benchmark built from real-world observations of five frontier model trajectories across nine established benchmarks (Tool Decathlon, Terminal-Bench 1.0 & 2.0, OSWorld-Verified, and more). Every evaluation sample is paired with a ground-truth observation from real environment execution, scored across five dimensions: format, factuality, consistency, realism, and quality.

Benchmark Results: Beating GPT-5.4

Qwen-AgentWorld-397B-A17B achieves the highest overall score of 58.71, surpassing GPT-5.4 (58.25), Claude Opus 4.8 (56.87), and Gemini 3.1 Pro (56.12). The advantage is most pronounced on Terminal and SWE — precisely where accurate modeling of code execution state and tool API behavior matters most.

The smaller variant, Qwen-AgentWorld-35B-A3B (a Mixture-of-Experts architecture with 35B total and 3B active parameters), scores 56.39 — placing it above Claude Sonnet 4.6 (56.04). The three-stage pipeline alone lifted its score by +8.66 points from baseline.

Inside the World Model's Mind

Analysis of 129 thinking traces reveals three emergent reasoning patterns that go beyond surface-level prediction:

Deliberative self-correction. The model uses "Wait!" as a cognitive interrupt to revise intermediate predictions — 1,347 such interrupts across 129 turns (10.4 per turn), covering factual errors and epistemological limits ("I cannot actually execute np.random.seed(42)").

Information leakage prevention. In the Search domain, the model holds a reference answer the agent is trying to find. When queries are unrelated, it prevents snippets from accidentally revealing the target — effectively a theory-of-mind equivalent for world models.

Multi-step causal reasoning. Predicting the output of curl -s localhost:3000 | python3 -m json.tool requires a six-step causal chain: Node.js missing → server never started → no listener → curl fails silently → empty pipe → JSONDecodeError.

Two Paradigms for Agent Enhancement

1. Decoupled Simulator (Sim RL)

As a standalone environment simulator, Qwen-AgentWorld replaces real environments during agent RL training. Key findings:

• Zero-shot generalization: Simulating 4,000 OpenClaw environments completely absent from training yields gains of +4.3 on Claw-Eval and +7.1 on QwenClawBench — with no domain-specific adaptation.
• Controllability is everything: Standard Sim RL without control instructions provides negligible gains. With targeted perturbations (intermittent API errors, paginated responses, incomplete intermediate results), MCPMark jumps +12.3 and WideSearch +16.3.
• Surpassing real environments: Controllable Sim RL achieves 50.3% F1 vs. 45.6% for Real RL trained against a live search engine. Sim-trained agents also learn different, more effective behaviors — they use full-page extraction more because simulated snippets deliberately withhold complete content.

2. Agent Foundation Model (LWM Warm-Up)

When world modeling is internalized directly into the agent, single-turn LWM RL training (with no tool calls) transfers to multi-turn agent tasks across seven benchmarks — including three entirely out-of-domain:

• Claw-Eval: +11.3
• QwenClawBench: +9.7
• BFCL v4: +9.0

The training pipeline contained zero Claw or function-calling data — these are genuinely transferred capabilities, not domain-specific shortcuts.

Data Sovereignty and the European Angle

For European organizations, Qwen-AgentWorld's open-source nature is significant. The 35B-A3B variant is released under a permissive license on Hugging Face and GitHub. With only 3B active parameters (MoE architecture), it runs on a single GPU with ~24 GB VRAM — such as an RTX 4090 or an A10G instance — making self-hosted deployment practical even for mid-sized European companies.

This matters under the EU AI Act and GDPR: organizations can deploy AgentWorld on their own infrastructure, keeping agent training data within EU boundaries. The model supports SGLang and vLLM with a 256K context window, and the evaluation pipeline (AgentWorldBench) uses OpenAI-compatible APIs that work with any self-hosted endpoint.

While Qwen-AgentWorld itself has no specific Czech, Slovak, or other European language optimization, its open architecture means fine-tuning for European languages and domain-specific environments (manufacturing, logistics, public administration) is feasible. The controllable simulation paradigm — where the world model can be instructed to simulate specific operational scenarios — is particularly relevant for European industries with complex compliance requirements.

What This Means for the Future of AI Agents

Qwen-AgentWorld represents a conceptual shift. Rather than scaling agents with ever-larger general-purpose models and hoping they figure out environments along the way, the Qwen team trained a model to explicitly understand environments. The implications extend well beyond academic research:

• Safer agents: An agent that mentally simulates consequences before acting is less prone to catastrophic errors in production.
• Scalable training: Thousands of simulated environments without dedicated infrastructure mean faster, cheaper agent development cycles.
• Systematic weakness exposure: Targeted perturbations reveal agent failure modes that real environments may never trigger — but production eventually will.
• Foundation for reasoning: Next-state prediction as internalized reasoning — "predict before you act" — may become a standard component of future agent architectures.

The research paper is available at arXiv:2606.24597. The 35B-A3B model, AgentWorldBench, and evaluation code are all open-source — and already deployable today.