Qwen just released Qwen-AgentWorld-35B-A3B, the first "language world model" designed to simulate agentic environments across seven domains. Released June 24, 2026, it takes a fundamentally different approach to agent training: instead of building better agents, Qwen built a model that simulates the environments agents operate in.
What Is a Language World Model?
A world model predicts what happens next in an environment given a current state and an action. Think of it like a physics engine, but for digital tasks instead of physical objects.
Qwen-AgentWorld simulates what a terminal returns when you run a command, what a search engine returns when you query it, what an app does when you tap a button. It predicts the next environment state given an agent's action and interaction history, using long chain-of-thought reasoning.
This matters because you can train agents inside simulated environments at scale, without consuming real API calls or risking real systems. If the simulation is good enough, agents trained in it generalize to real environments.
Seven Domains, One Model
Previous world models handled single domains. Qwen-AgentWorld covers seven in a single model:
- MCP (Model Context Protocol tool calling)
- Search (web search interaction)
- Terminal (Linux command line)
- SWE (software engineering tasks)
- Android (mobile UI interaction)
- Web (browser interaction)
- OS (desktop operating system)
This is a significant scope expansion. An agent trainer can use one model to simulate everything from API calls to mobile apps to web browsing.
Architecture and Training
Specs at a glance:
| Attribute | Value |
|---|---|
| Total parameters | 35B |
| Active parameters | 3B (MoE) |
| Context length | 262,144 tokens (256K) |
| Experts | 256 (8 routed + 1 shared) |
| Base model | Qwen3.5-35B-A3B-Base |
| Architecture | Gated DeltaNet + Gated Attention hybrid |
The training pipeline has three stages:
- Continual Pre-Training (CPT): Injects environment knowledge into the base model using 10M+ real-world interaction trajectories.
- Supervised Fine-Tuning (SFT): Activates next-state-prediction reasoning, teaching the model to predict what an environment returns.
- Reinforcement Learning (RL via GSPO): Sharpens simulation fidelity through reward signals.
The key insight: environment modeling is the training objective from stage one onward. This is not a general-purpose LLM fine-tuned to simulate environments. It was built for this purpose from the ground up.
Benchmark Results: Competitive With Frontier Models
Qwen also released AgentWorldBench, evaluating models on five dimensions per domain: Format, Factuality, Consistency, Realism, and a fifth rubric.
Overall scores (higher is better):
| Model | Overall |
|---|---|
| Qwen-AgentWorld-397B-A17B | 58.71 |
| GPT-5.4 | 58.25 |
| Claude Opus 4.6 | 57.80 |
| Claude Opus 4.8 | 56.59 |
| Qwen-AgentWorld-35B-A3B | 56.39 |
| Claude Sonnet 4.6 | 56.04 |
| Gemini 3.1 Pro | 54.57 |
| Qwen3.5-397B-A17B | 54.74 |
| Kimi K2.6 | 53.42 |
| DeepSeek-V4-Pro | 52.97 |
| GLM-5.1 | 51.31 |
The larger 397B version takes the top spot, narrowly beating GPT-5.4. The 35B open-weight model scores 56.39 overall, beating DeepSeek-V4-Pro (862B), Gemini 3.1 Pro, and Kimi K2.6.
The 35B model shows an 8.66-point improvement over its base model (Qwen3.5-35B-A3B at 47.73), directly attributable to the world model training pipeline.
The Interesting Part: Agents Trained in Simulation Generalize to Reality
The most striking finding is that agents trained inside Qwen-AgentWorld's simulation generalize to real, out-of-distribution environments.
On the Claw-Eval benchmark (an agent framework not seen during training), agents trained with Qwen-AgentWorld-397B-A17B simulation scored 69.7, compared to 65.4 for the base model. That is a 4.3-point gain from training in a simulated environment that the world model had never encountered before.
Even more counterintuitively, agents trained in fully fictional, self-consistent worlds (controlled simulation) generalized better to real search tasks than agents trained on real data. The fictional-world search F1 improved by 16.29 points for the 35B model. Qwen hypothesizes that fictional worlds provide more diverse training signal than the narrow distribution of real queries.
Agent Foundation Model Transfer
The RL warm-up used for world model training also improves the model's performance as a general agent. Single-turn RL warm-up transferred to multi-turn tool-calling tasks across seven benchmarks, including three entirely out-of-domain:
- Terminal-Bench 2.0: +6.3 points
- SWE-Bench Verified: +3.4 points
- BFCL v4: +9.0 points
- WideSearch F1: +12.8 points
This suggests world modeling and agent capabilities are complementary skills that reinforce each other.
Why This Matters
Most AI agent research focuses on building better agents (better reasoning, better tool use, better planning). Qwen-AgentWorld flips this: build a better environment simulator, then train agents inside it.
If this approach scales, it could change how AI agents are developed. Instead of learning through expensive real-world interactions (API calls, browser sessions, terminal commands), agents could train at scale inside high-fidelity simulations. The simulation handles the environment, the agent handles the task.
The fact that the 35B model is open-weight makes this immediately accessible to researchers and developers who want to experiment with simulated agent training.
Availability
- Model weights: Qwen/Qwen-AgentWorld-35B-A3B on HuggingFace
- Benchmark: AgentWorldBench dataset on HuggingFace
- Code: QwenLM/Qwen-AgentWorld on GitHub
- Technical report: arXiv:2606.24597
- License: Check HuggingFace model card
- Serving: Compatible with vLLM, SGLang, and HuggingFace Transformers
Requires at least 4 GPUs for the 35B model with 256K context. Qwen recommends maintaining at least 128K context for multi-turn simulation fidelity.