StepOpen Releases 4B Agent Model — Runs on All Android Devices, One-Click Deployment for DIY Enthusiasts
GELab-Zero — First Simultaneous Open-Source GUI Agent Model with Full Infrastructure & One-Click Deployment
StepStar (阶跃星辰) has officially open-sourced GELab-Zero, a powerful GUI Agent model and complete deployment infrastructure.
Performance Highlights
- GELab-Zero-4B-preview has set new records on multiple mobile and desktop GUI leaderboards for its size category.
- Achieves State-of-the-Art (SOTA) results in GUI understanding and execution.
- Outperforms larger models (e.g., GUI-Owl-32B) while delivering better deployment agility.
---
Why GELab-Zero Matters
With AI now prevalent in consumer devices like smartphones, Mobile Agents are transitioning from possible to scalable.
GUI Agents offer:
- Adaptation to almost any app based purely on visual understanding.
- No vendor-specific changes needed, resulting in minimal integration cost.
StepStar also released AndroidDaily, a real-world evaluation standard for consumer-grade GUI model capability testing.
---
The Challenge in GUI Agent Development
Running a mobile GUI Agent across devices and OS variations is hard due to:
- Fragmented ecosystems.
- Complex setups: multi-device ADB, dependency installation, permissions, inference service handling, orchestration, and replay.
Solution: Lower the entry barrier so developers can focus on innovation without reinventing the core infrastructure.
---
What's Included in GELab-Zero
- Local GUI Agent Model — GELab-Zero-4B-preview
- Plug-and-Play Full Inference Infrastructure — handles heavy engineering tasks automatically.
- AndroidDaily Benchmark Suite — designed around genuine business scenarios.
Benchmarks Used:
- ScreenSpot
- OSWorld
- MMBench
- Android World
These cover:
- GUI comprehension
- Element locating
- Interactive operations
Result: GELab-Zero-4B-preview delivers SOTA performance in its category.
---
Example Scenarios
Scenario 1: Multi-Item Purchase in Food Delivery Apps
Prompt: Buy multiple items of specified categories, specs, and quantities via Ele.me’s nearest Hema Fresh store.
Outcome: Model accurately identifies items and executes repetitive multi-step purchases smoothly.
---
Scenario 2: Claim Enterprise Meal Voucher
Prompt: Navigate within GeiDao to claim a specific “Employee Benefits” voucher.
Outcome: Handles niche app navigation with precision.
---
Scenario 3: Play a Classic Movie Featuring a Specific Actor
Prompt: On Tencent Video, play a classic Jackie Chan action film.
Outcome:
- Recognizes subjective term (“classic”).
- Closes pop-ups.
- Searches within the movie category and selects top-rated film.
---
Scenario 4: Weekend Activity for Kids
Prompt: Find an activity spot for a child in Beijing.
Outcome:
- Searches content platforms.
- Evaluates options & recommends “Wanku Adventure” at Beijing Garden Expo Park.
- Highlights kid-friendly features.
---
Key Capabilities
- Lightweight Local Inference
- Run 4B models on consumer-grade hardware with low latency & privacy protection.
- One-Click Task Initiation
- Unified deployment handles dependencies & device setup automatically.
- Multi-Device Task Distribution
- Assign tasks across multiple devices and record interaction traces.
- Multiple Agent Modes
- Supports ReAct closed-loop, multi-agent collaboration, and scheduled tasks.
---
AndroidDaily Benchmark
Developed to reflect real-world everyday tasks — beyond productivity apps — covering:
- Food & Dining
- Travel
- Shopping
- Housing
- Information Consumption
- Entertainment
Accuracy:
GELab-Zero-4B-preview scored 73.4% on AndroidDaily across complex mobile scenarios.
---
Dual-Track Evaluation System
1. Static Evaluation
- Tests grounding (UI understanding) & action planning.
- Dataset: 3,146 actions with step-by-step screenshots.
- Measures numeric prediction accuracy (e.g., clicks, input).
2. End-to-End Testing
- Full-task execution in real or emulated environments.
- Covers:
- Transportation (ride-hailing, navigation, public transit)
- Shopping & Payment
- Social Communication
- Content Consumption
- Local Services
- Metric: Overall success rate per scenario.
---
Open-Source Access
- GitHub: https://github.com/stepfun-ai/gelab-zero
- Hugging Face: https://huggingface.co/stepfun-ai/GELab-Zero-4B-preview
---
Monetization & Ecosystem Integration
For creators and developers, AiToEarn complements GELab-Zero by enabling AI-powered content monetization and cross-platform publishing.
Supports simultaneous distribution to:
- Douyin, Kwai, WeChat, Bilibili, Rednote
- Facebook, Instagram, LinkedIn, Threads, YouTube, Pinterest, X (Twitter)
Integrated features:
- AI generation tools
- Cross-platform analytics
- Model ranking (AI模型排名)
Bottom line: GELab-Zero lowers the deployment barrier for GUI Agents, while tools like AiToEarn empower creators to monetize innovations at scale.
---
Would you like me to add a quick “Getting Started” section with installation and run commands for GELab-Zero so developers can use this right after reading? That could make this Markdown more actionable.
