GELab-Zero 4B Preview
GELab-Zero 4B Preview is StepFun's GUI agent release based on Qwen3-VL-4B-Instruct. It uses screen images to select actions across mobile and computer interfaces.
Preview
The model needs an action runner
The GELab-Zero project includes separate code for ADB connections, dependency setup and task recording or replay. These functions belong to the surrounding system; the weights alone do not connect to a phone.StepFun GELab-Zero-4B-preview model card
Check each action against the next screen
Practical guidance. A valid click prediction is only one step in a task. The application must obtain a fresh screen after each action and detect whether the expected change occurred. Start with short, reversible workflows before longer sequences.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Status | Preview |
| Access | Downloadable |
| Model identifiers | GELab-Zero-4B-preview |
| Ancestry | Qwen/Qwen3-VL-4B-Instruct is the declared base model.StepFun GELab-Zero-4B-preview model card |
| License | Apache 2.0.StepFun GELab-Zero-4B-preview model card |
| Actions | The card lists click, type, slide and wait actions.StepFun GELab-Zero-4B-preview model card |
Related models
- Step3-VL-10B
StepFun ยท Step 3
Step3-VL-10B combines a 1.8B PE-lang visual encoder with a Qwen3-8B decoder.
Launch date not confirmedDownloadable