Overview
WebWorld-8B is a web world model from the Qwen team, fine-tuned from Qwen3-8B. It does not chat. Given the current state of a web page and an action, it predicts what the page becomes next — so it acts as a browser you can run in your head, in software, instead of clicking through the real site.
That solves a specific problem in agent training. Teaching a web agent on the live internet is slow and risky: pages rate-limit you, network latency stacks up, and a wrong click can submit a form or spend money. WebWorld replaces the live web with a simulator. The model was trained on more than a million real browser trajectories, and it keeps the page format you give it — A11y Tree, HTML, XML, Markdown, or plain language — across multi-turn rollouts of 30 or more steps.
What it is good at
WebWorld-8B is built for agent developers rather than end users. The common uses are:
- Lookahead search — an agent proposes several candidate actions, WebWorld simulates the resulting page for each, a value model scores them, and only the best action runs on the real browser.
- Training-data synthesis — it turns a handful of seed tasks into thousands of fine-tuning trajectories, keeping only the runs that succeed.
- Offline evaluation — you can replay and test agent behavior without touching production sites.
Fine-tuning Qwen3-8B on WebWorld-generated trajectories has been reported to add +9.9% on MiniWob++ and +10.9% on WebArena. The 8B size is the entry point of the series; 14B and 32B variants trade more hardware for higher fidelity.
Running it locally
WebWorld-8B is an 8B model, so it fits on consumer hardware — roughly 16–24 GB of VRAM at full precision, and a 4-bit quantized build runs on a 16 GB laptop. Pull the weights from Hugging Face and load them through your preferred runtime:
huggingface-cli download Qwen/WebWorld-8B
From there you can serve it with Transformers or vLLM, or load it in a desktop app like Atomic Chat, where it downloads and runs in one step with nothing leaving your device.
License
WebWorld is released under Apache 2.0 — the weights, the dataset, and the demo code. That means free local use, fine-tuning, and commercial deployment. The research behind it is documented in the WebWorld paper.