Qwen/WebWorld-8B

Hugging Face
TEXT GENERATIONConcurrency Cost:1Model Size:8BQuant:FP8Ctx Length:32kPublished:Feb 13, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Warm

Qwen/WebWorld-8B is an 8 billion parameter open-web world model developed by Qwen, built upon Qwen3-8B, designed for training and evaluating web agents. It is trained on over 1 million real-world web interaction trajectories, enabling long-horizon simulation (30+ steps) and supporting multi-format state representations like A11y Tree, HTML, and natural language. This model excels at predicting web page states and agent behavior, outperforming GPT-5 as a world model for inference-time lookahead search.

Loading preview...

WebWorld-8B: A Specialized Web World Model

Qwen/WebWorld-8B is an 8 billion parameter model from the WebWorld series, developed by Qwen, specifically engineered as an open-web world model for training and evaluating web agents. Built on the Qwen3-8B architecture, it leverages over 1 million real-world web interaction trajectories, processed through a scalable hierarchical data pipeline.

Key Capabilities

  • Long-horizon simulation: Supports complex web interactions spanning 30+ steps.
  • Multi-format state representations: Understands and generates web states in various formats including A11y Tree, HTML, XML, Markdown, and natural language.
  • CoT-activated reasoning: Incorporates Chain-of-Thought reasoning for enhanced transition prediction.
  • Cross-domain generalization: Demonstrates strong performance across code, GUI, and game environments.
  • Superior performance: Agents trained with WebWorld-synthesized trajectories achieve significant gains (+9.9% on MiniWob++ and +10.9% on WebArena). For inference-time lookahead search, WebWorld has been shown to outperform GPT-5 as a world model.

Good For

  • Web agent development: Ideal for training and evaluating autonomous web agents.
  • Fast simulation: The 8B version is recommended for rapid simulation and data synthesis tasks.
  • Predicting web state changes: Accurately predicts the next page state given the current state and an action, preserving input/output formats.
  • Multi-turn trajectory simulation: Capable of simulating complex user journeys over many steps.

Finetunes
1 models