WebWorld

model
Alibaba

Open web world model for training and evaluating browser agents: predicts next-page state given the current state and a candidate action. Three sizes (8B / 14B / 32B) fine-tuned on Qwen3 bases over 1M+ real-world web interaction trajectories, supporting five state representations (A11y Tree, HTML, XML, Markdown, natural language). Supports long-horizon (30+ step) trajectory simulation with CoT-activated transition prediction.

Intrinsic (WebWorld-Bench): Factuality 71.0, Web Turing 45.6. When used for agent training, downstream gains: +9.9% on MiniWob++, +10.9% on WebArena. Cross-domain generalization to GUI desktop (+0.383), code (+0.249), and game environments (+0.220) — the world model abstraction transfers beyond browser tasks. Apache 2.0.

Model weights re-released on HuggingFace May 8, 2026.

Paper (arXiv)GitHubHuggingFace collectionDataset

Model Details

Architecture DENSE
Parameters 32B
License Apache 2.0
Base model qwen3

Benchmark Scores

Benchmark Score Mode
WebWorld-Bench Factuality 71.0 —
WebWorld-Bench Web Turing 45.6 —

Variants

Name Parameters Notes
WebWorld-8B 8B Fast simulation tier; built on Qwen3-8B
WebWorld-14B 14B Built on Qwen3-14B
WebWorld-32B 32B Highest fidelity; built on Qwen3-32B

Paper

arXiv HTML
agenticbrowser-agentopen-weightfoundational

Related