What is WebWorld-8B?
WebWorld-8B is a web world model from the Qwen team, a fine-tune of Qwen3-8B that simulates the web instead of browsing it. Given the current page state and an action, it predicts the full next page state, and it keeps that up across multi-turn trajectories of 30+ steps. The team trained the series on more than 1M real-world web interaction trajectories collected through a hierarchical data pipeline, and lists three sizes in it: 8B, 14B and 32B. The 8B repository went up on February 13, 2026 under Apache 2.0, and Qwen's own recommendation is to use the 8B for fast simulation and data synthesis, which is also the role that fits local hardware best: the largest GGUF build in the table below is 6.73 GB.
| Specification | WebWorld-8B |
|---|---|
| Total parameters | 8.2B |
| Base model | Qwen3-8B |
| Model type | Web world model: predicts the next page state from the current state and an action |
| Training data | 1M+ real-world web interaction trajectories (WebWorldData) |
| Page state formats | A11y Tree, HTML, XML, Markdown, natural language |
| Context window | Not stated on the model card, which tags the model long-context |
| Simulation horizon | Multi-turn, 30+ steps |
| Modalities | Text in, text out |
| Languages | English, Chinese |
| Series | 8B, 14B, 32B |
| Release date | February 13, 2026 |
| License | Apache 2.0 |
The model strictly preserves whatever state format you feed it: an accessibility tree in means an accessibility tree out, and the same holds for HTML, XML, Markdown and plain language. Actions arrive as Python-style function calls covering clicks, form fills, scrolling, navigation and tab management, and transition prediction runs on chain-of-thought reasoning. The model card is upfront about three limits: outcomes can skew optimistic in favour of the agent's intended action, long-form high-precision content is not the target, and the model is text-only, so it does not simulate pixel-level rendering.
WebWorld-8B benchmarks
The numbers below are the Qwen team's own, from the model card: WebWorld-Bench scores world models on Factuality (functional correctness) and Web Turing (perceptual realism) across nine dimensions, and the cross-domain rows measure simulation quality outside the web against the base model:
| Benchmark | WebWorld-8B | Qwen3-8B (base) | GPT-4o | Gemini-3-Pro | Claude Opus 4.1 |
|---|---|---|---|---|---|
WebWorld-Bench Factuality Functional correctness | 70.1 | 26.9 | 59.5 | 70.3 | 71.3 |
WebWorld-Bench Turing Perceptual realism | 42.2 | 17.4 | 35.4 | 43.2 | 47.4 |
GUI Desktop Desktop simulation | 0.705 | 0.322 | - | - | - |
Code Code simulation | 0.396 | 0.147 | - | - | - |
Game Game simulation | 0.473 | 0.253 | - | - | - |
API Services API simulation | 0.299 | 0.088 | - | - | - |
The fine-tune moves Qwen3-8B from 26.9 to 70.1 on factuality, within 0.2 points of Gemini-3-Pro, while Claude Opus 4.1 keeps the lead on both WebWorld-Bench scores. The result agent builders care about sits outside this table: agents fine-tuned on WebWorld-synthesized trajectories gain +9.9% on MiniWob++ and +10.9% on WebArena, and the team reports that WebWorld beats GPT-5 as a world model for inference-time lookahead search.
WebWorld-8B hardware requirements
The system requirement to check is memory. The community imatrix quants at mradermacher/WebWorld-8B-i1-GGUF cover the whole ladder; these are the real file sizes:
| Memory | Build to pick | File size |
|---|---|---|
| 4 GB | i1-IQ2_M | 3.05 GB |
| 6 GB | i1-Q3_K_M | 4.12 GB |
| 8 GB | i1-Q4_K_M | 5.03 GB |
| 12 GB | i1-Q5_K_M | 5.85 GB |
| 16 GB and up | i1-Q6_K | 6.73 GB |
Neighbouring builds in this table sit 0.8 GB to 1.1 GB apart, so when two builds both fit, take the larger one. If the format is new to you, start with what GGUF is.
How to run WebWorld-8B in Atomic Chat
Atomic Chat is a free local app for macOS, Windows and Linux. It includes a Hugging Face model browser and a built-in chat, with no manual llama.cpp build required.
- Download Atomic Chat for your platform and open it.
- Search for WebWorld-8B in the model browser and open Download Options.
- Pick the build that fits the memory you have, then start a chat.
For the rest of the family, see every Qwen model you can run locally, or the base model this one was fine-tuned from, Qwen3-8B.
WebWorld-8B license
WebWorld-8B is released under Apache 2.0. That permits commercial use, modification and redistribution with no royalties, so you can run simulations, synthesize training data and ship products on top of the model without a usage fee. The WebWorldData training dataset and the WebWorld code on GitHub are published alongside the weights.