Training an AI agent to use a terminal or browser correctly is awkward in a way that training a chat model is not. A chat model can be trained on a huge dataset of existing text; the environment just sits there. An agent has to act, get a response from the environment, decide what to do next, and eventually reach a goal. That loop requires a live environment — which means real execution, real state, real latency, and real failure modes. Running a million RL episodes against actual Linux shells or web browsers is not fast.
The obvious move is to replace the real environment with a simulator. If you can build something that behaves enough like a bash shell that an agent policy trained against it generalizes to the real one, you’ve bought yourself cheaper and more parallelizable training. The problem has always been building the simulator: physical robotics research has used learned world models for years (Dreamer-style approaches, for instance), but the digital domain is different — the “physics” of an OS terminal are exactly the rules of Bash, and those rules are already encoded in a large language model’s weights.
Qwen-AgentWorld, released by Alibaba’s Qwen team on June 23, makes this explicit. Instead of building a rule-based simulator, they trained an LLM to predict what happens next when an agent takes an action in one of seven digital environments: MCP, Search, Terminal, Software Engineering, Android, Web, and OS. The result is a language world model — a model whose output is the next environment state and observation, rather than a helpful response.
Two models are released as open weights under Apache 2.0: Qwen-AgentWorld-35B-A3B (35B total parameters, 3B active per token) and Qwen-AgentWorld-397B-A17B (397B total, 17B active). Both use a sparse mixture-of-experts architecture and a 256K context window. Training went through three stages: general pretraining to inject world knowledge and state-transition dynamics, supervised fine-tuning to activate “next-state prediction thinking” patterns, and RL refinement with hybrid rubric-and-rule rewards to sharpen simulation fidelity. The training corpus was over 10 million environment-interaction trajectories, collected from five frontier models — including Claude Opus 4.6 — running against nine established agent benchmarks.
The evaluation benchmark they built, AgentWorldBench, scores predicted next states across five dimensions: Format, Factuality, Consistency, Realism, and Quality. The 397B variant reportedly outperforms all frontier proprietary models on overall scoring, though taking self-reported agent benchmarks at face value requires the usual caution. More practically interesting is what the models are supposed to enable: they demonstrate two downstream applications — using the world model as a standalone simulator to train downstream RL policies more cheaply, and using it as a warm-start foundation model whose pre-trained dynamics knowledge improves performance across seven agent benchmarks when fine-tuned.
The GUI domains are worth noting: for Android, Web, and OS environments, observations are represented as accessibility trees and UI view hierarchies rather than pixel frames. That’s a practical choice (representing a full screenshot as tokens is expensive and lossy), but it also means the world model is operating on the structured representation of UI state rather than learning raw visual dynamics. Whether that limits transfer to agents that operate on visual observations is an open question.
The comparison to Qwen-RobotWorld, released a week earlier, is instructive. That model predicts future physical states from observations and natural-language action descriptions, serving as a look-ahead tool for robotic navigation and manipulation policies. AgentWorld does the same thing but for digital action spaces — the “physics” being the rules of file systems, browser DOMs, and terminal state machines rather than rigid-body dynamics. Both represent the same underlying bet: that a learned simulator, trained at scale on interaction data, is tractable enough to be useful for downstream agent training. The digital domain is arguably more tractable because the ground truth is deterministic (a shell command has a correct output) rather than physically continuous.
One thing that’s hard to evaluate from the paper alone is whether the world model’s fidelity is actually high enough to produce policies that transfer to real environments. A world model that gets 90% of shell commands right will produce agents that fail reliably in the 10% of cases where the simulator diverged from reality — and those cases may not be uniformly distributed. The team shows gains on seven agentic benchmarks, but those benchmarks are finite, and whether the gains generalize to open-ended real-world tasks will depend on how well the simulation fidelity distributes across the long tail of environment behaviors.
The code and weights are available on GitHub. For anyone working on agent training infrastructure, the 35B-A3B model in particular seems like a reasonable thing to evaluate — the active-parameter count keeps inference manageable, and having a model that can stand in for a live environment during RL rollouts is a meaningful practical primitive even if it’s not perfect.