GUI-agent researchers
Reproduce grounding and mobile-navigation experiments with official code, checkpoints, evaluation material and a technical report.
Independent tool overview
MAI-UI at a glance
MAI-UI is Alibaba's research-oriented family of vision-language models for understanding and operating mobile interfaces. The open release provides 2B and 8B weights plus agent code for GUI grounding and Android navigation. It is a developer foundation—not a consumer phone assistant—and Alibaba now identifies Qwen-UI-Agent as its next-generation successor.
Visit the official MAI-UI site ↗
- Best for
- Mobile GUI-agent research and prototypes
- Public weights
- MAI-UI-2B and MAI-UI-8B
- Model family
- 2B, 8B, 32B and 235B-A22B
- Core actions
- Click, type, swipe, drag, wait and system controls
- Extended actions
- Ask-user and MCP tool calls
- Successor
- Qwen-UI-Agent
Overview
What MAI-UI is
MAI-UI takes screenshots and natural-language instructions, identifies interface elements and returns actions such as click, swipe, type, long-press, drag, wait or press a system button. Its extended action space can also ask the user for missing information and call MCP tools when an API is more reliable than a long sequence of taps.
The research family spans 2B, 8B, 32B and 235B-A22B variants, but the public quick start links downloadable weights only for MAI-UI-2B and MAI-UI-8B. Developers serve a model through vLLM, run the supplied grounding or navigation notebooks and connect it to a controlled device environment.
MAI-UI remains useful for reproducing mobile-agent research and building experiments, but it is no longer the team's leading model. In July 2026, the same repository introduced Qwen-UI-Agent, extending the work across phones, computers, browsers and DeepSearch. New production-oriented evaluations should compare the successor before standardizing on MAI-UI.
Use cases
Who MAI-UI is best for
The strongest fit depends on the job you need the product to complete, not the size of its feature list.
Android automation prototypes
Test screenshot-driven task execution across apps when normal APIs are unavailable or incomplete.
On-device model exploration
Evaluate the 2B checkpoint and the paper's device-cloud routing ideas under explicit hardware and privacy constraints.
MCP-augmented agent experiments
Study an action space that can switch between interface control, user clarification and structured external tools.
Capabilities
Core MAI-UI features
Visual GUI grounding
Locate interface elements from a screenshot and a natural-language instruction without requiring an accessibility tree.
Multi-step mobile navigation
Generate sequences of taps, swipes, typing and system actions for tasks that cross screens or applications.
Agent-user interaction
Use an ask-user action to request missing information or confirmation instead of guessing silently.
MCP tool augmentation
Call structured tools when an API can replace brittle or inaccessible interface operations.
Device-cloud collaboration research
Route work between a smaller local model and a more capable cloud model based on state and data sensitivity.
Open code and checkpoints
Run the Apache-2.0 repository and download the public 2B or 8B model weights from Hugging Face.
Process
How the MAI-UI workflow works
Step 1
Choose MAI-UI or its successor
Use MAI-UI for reproduction and mobile-only work; compare Qwen-UI-Agent for a newer cross-device project.
Step 2
Build an isolated test environment
Use emulators or dedicated test phones with synthetic accounts, restricted permissions and no personal payment or message data.
Step 3
Serve a public checkpoint
Deploy MAI-UI-2B or 8B through the repository's supported vLLM configuration and expose the local OpenAI-compatible endpoint.
Step 4
Run the official cookbook
Validate grounding first, then navigation, before connecting the model to a broader action executor.
Step 5
Add hard action policies
Require explicit approval for purchases, messages, account changes, destructive actions and any step involving secrets.
Step 6
Measure real task reliability
Test across app versions, screen sizes, pop-ups, network failures and ambiguous instructions instead of relying only on benchmark scores.
Cost
MAI-UI pricing and free plan
MAI-UI is an open-source research release rather than a hosted subscription. The repository and public 2B and 8B checkpoints can be downloaded without a license fee, but users pay for their own GPU inference, device lab, engineering, storage and any cloud or MCP services. The project does not publish a managed commercial price.
Open-source code
Free
Alibaba's MAI-UI agent implementation, cookbooks and evaluation code.
- Apache License 2.0 for the MAI-UI Mobile repository
- Grounding and navigation notebooks
- Local OpenAI-compatible model endpoint through vLLM
- Third-party components retain their own licenses
MAI-UI-2B
Free weights; infrastructure extra
The smaller public checkpoint for constrained and on-device-oriented experiments.
- Official Hugging Face checkpoint
- Lower compute requirement than the 8B model
- Intended for GUI grounding and mobile navigation
- Hardware compatibility and performance must be tested by the implementer
MAI-UI-8B
Free weights; infrastructure extra
The larger public checkpoint used by the official quick-start examples.
- Official Hugging Face checkpoint
- Served with the documented vLLM configuration
- Grounding and full navigation examples
- GPU, device and operating costs are not included
32B and 235B-A22B variants
No public hosted price
Research variants described and benchmarked in the technical report.
- Part of the reported MAI-UI model family
- Used for higher-capability benchmark results
- Not linked as downloadable checkpoints in the public quick start
- Do not assume public availability from the benchmark description alone
Pricing checked . Check current pricing at the source ↗
Assessment
MAI-UI strengths and limitations
Where it stands out
- Public 2B and 8B checkpoints support hands-on reproduction rather than paper-only evaluation.
- Covers both GUI grounding and multi-step mobile navigation.
- Ask-user actions let the agent surface ambiguity instead of always acting autonomously.
- MCP integration can replace long, brittle interface sequences with structured tools.
- The small-to-large research family examines realistic device and cloud tradeoffs.
- Official notebooks provide concrete starting points for grounding and navigation.
- Apache-2.0 repository licensing is friendly to experimentation and integration.
What to consider
- MAI-UI is a developer research project, not an installable consumer assistant or managed API.
- Only the 2B and 8B weights are linked in the public quick start; larger benchmarked variants should not be treated as generally available.
- The documented vLLM and model-serving stack requires technical setup and suitable compute.
- GUI agents remain vulnerable to layout changes, pop-ups, timing issues and cascading action errors.
- Benchmark success does not establish safe or reliable behavior on a user's own applications and accounts.
- Autonomous clicks and typing can send messages, change data or initiate purchases, so external approval controls are essential.
- The public demos are heavily oriented toward Chinese mobile apps and may not transfer directly to other regions or app ecosystems.
- Qwen-UI-Agent is now the maintainers' newer cross-platform direction, reducing the case for starting a greenfield project on MAI-UI alone.
Compare
MAI-UI alternatives
The right alternative depends on the specific output, workflow, controls and budget your project requires.
Agents
Gemini Computer Use
A managed computer-use model for teams that prefer a hosted API over self-serving mobile GUI checkpoints.
Explore Gemini Computer Use →Consumer
ChatGPT Atlas
A user-facing AI browser for web tasks when mobile-app control and model research are not required.
Explore ChatGPT Atlas →Agents
Manus
A broader hosted agent for delegated research and productivity work without building a custom device-control stack.
Explore Manus →Questions
MAI-UI FAQs
What is MAI-UI?
MAI-UI is Alibaba's family of vision-language GUI agent models for interface grounding and multi-step mobile navigation.
Can MAI-UI control a smartphone?
Yes, when developers connect the model to an action executor and device environment. It can propose clicks, swipes, typing, drags and system-button actions from screenshots.
Is MAI-UI a phone app?
No. It is an open research repository and model release that developers must deploy and integrate themselves.
Which MAI-UI models can I download?
The official quick start links public Hugging Face checkpoints for MAI-UI-2B and MAI-UI-8B.
Are the 32B and 235B MAI-UI models public?
They are described and benchmarked in the technical report, but the public quick start does not link downloadable checkpoints for them.
How much does MAI-UI cost?
The code and public checkpoints have no subscription price. Users supply and pay for their own inference hardware, device environment, engineering and external services.
What does MCP add to MAI-UI?
MCP lets the agent call structured external tools when that is more reliable or capable than manipulating every step through the phone interface.
Is MAI-UI open source?
The MAI-UI Mobile repository states that it is licensed under Apache License 2.0, with separate licenses applying to included third-party components.
What replaced MAI-UI?
The maintainers describe Qwen-UI-Agent as the follow-up generation, extending the work to mobile, desktop, browser and DeepSearch scenarios.
Bottom line
Our MAI-UI verdict
MAI-UI is a useful open baseline for teams studying mobile GUI grounding, navigation, user clarification and MCP-augmented actions. Its public checkpoints and notebooks make it practical for experiments, but it is not a turnkey automation product and should never be connected directly to valuable accounts without hard controls. For a new cross-device research program, evaluate the newer Qwen-UI-Agent alongside it.
Visit MAI-UI website ↗Similar Agents Tools
View all Agents AI tools →
Claude in Chrome
Claude in Chrome - Agentic browser extension to use Anthropic's Claude directly in Google Chrome
Grok Voice Agent
Grok Voice Agent - Build voice agents that speak dozens of languages, call tools, and search realtime data.
Nemotron 3
Nemotron 3 - Nvidia's new family of open-source AI models in three sizes designed to help developers build multi-agent systems more efficiently
Claude Cowork
Claude Cowork - Anthropic's new MacOS tool that brings Claude Code's agentic capabilities to everyday tasks
AI training for the future of work.
Get access to all our AI courses, hundreds of real-world AI use cases, live expert-led workshops, an exclusive network of AI early adopters, and more.
AI Courses
Get unlimited access to all of our current & upcoming industry-specific AI courses for the duration of your subscription.
Daily Guides
To keep up with the rapid pace of AI, our team publishes AI implementation guides daily. Our library contains 300+ practical use cases to automate real-world work.
Workshops
Join weekly, live, interactive sessions with industry leaders who are at the forefront of AI for hands-on implementation guidance and exclusive insights.
Community
Network with an exclusive community of AI-first professionals who are working smarter with AI. Learn how early adopters are using AI in their work and businesses.