Alibaba's New AI Agent MAI-UI is Beating Google's Gemini on Android

Have you ever tried to get an AI assistant to do something on your phone and just given up? You ask it to "book the earliest flight to Austin next Tuesday," and it just opens a web search. It's frustrating, right? We have these incredibly powerful AI brains in the cloud, but getting them to actually use our phones like a human would is still a huge challenge.

Well, it looks like the team at Alibaba's Tongyi Lab has been thinking about this a lot. They just dropped a paper on a new system they're calling MAI-UI, and it's a pretty big deal.

This isn't just another chatbot. MAI-UI is a family of AI agents designed from the ground up to navigate the messy, unpredictable world of mobile app interfaces. And the early results are staggering—it's already outperforming major players like Google's Gemini 2.5 Pro on some key Android navigation tests. Let's break down what they're doing differently, because it’s genuinely cool.

So, What Exactly is MAI-UI?

At its heart, MAI-UI is a multimodal AI. That just means it can understand more than one type of input. You give it a command in plain English and show it a screenshot of your phone's screen, and it figures out what to do next.

Think of it like you're looking over a friend's shoulder and telling them where to tap. "Hey, tap on your profile picture, then go to settings, and find the billing section." MAI-UI is designed to do exactly that, outputting structured actions like "click this button" or "swipe up."

But here's where it gets interesting. The team at Alibaba realized that just blindly following instructions isn't enough. A truly helpful agent needs to be smarter. So, they baked in three things that most other agents ignore:

1. It can talk back: If your command is vague, MAI-UI can actually ask you for more information. No more guessing games.
2. It can use other tools: It’s not just limited to tapping and swiping. It can call on other mini-programs or APIs to get things done, like a real assistant would.
3. It respects your privacy: It uses a clever "device-cloud" system. Simple tasks and sensitive data stay on your phone, but for heavy-duty thinking, it can call on the massive AI models in the cloud. It’s the best of both worlds.

This is all built on their own Qwen3 family of models, with sizes ranging from a nimble 2 billion parameters all the way up to a massive 235 billion.

How It Learns to "See" Your Screen

One of the hardest jobs for an AI like this is something called "GUI grounding." It's a fancy term for a simple idea: connecting your words to the right pixels on the screen. When you say, "open my monthly billing settings," the AI has to figure out which of the dozens of elements on the screen is the correct one to tap.

This is where a lot of models stumble. MAI-UI takes a smarter approach. Instead of just learning a single description for a button, the training pipeline gives the AI multiple perspectives.

Imagine you're describing a button to a friend. You might say:

• "It's the blue button." (Appearance)
• "It's the one for saving your work." (Function)
• "It's in the top right corner." (Location)
• "It's what you press when you're done." (User Intent)

MAI-UI learns from all these different angles at once. This makes it way more resilient when your instructions aren't perfect. The model treats all these descriptions as evidence to pinpoint the exact spot to "tap."

And the results speak for themselves. On tough industry benchmarks, MAI-UI is setting new records, even surpassing models like Gemini 3 Pro and Seed1.8 on specific grounding tasks. It's a huge step toward an AI that truly understands what it's looking at.

The Secret to Navigating Complex Apps

Okay, so it can find the right button. But what about tasks that take multiple steps across different screens or even different apps? This is where navigation comes in, and it's a whole other level of difficulty.

To solve this, the Alibaba team created what they call a "self-evolving data pipeline." It sounds complex, but the idea is brilliant.

Think of it like this:

1. The Homework: They start with a set of basic tasks, like "order a coffee" or "check my messages."
2. The Practice Session: They have a team of AI agents (along with some human help) try to complete these tasks on real Android environments.
3. The Teacher's Red Pen: A "judge" model then watches these attempts. It identifies which ones were successful and filters out all the failed or low-quality attempts.
4. The Next Lesson Plan: The successful attempts are then added back into the training data for the next round.

It's a continuous feedback loop. The AI is constantly creating its own high-quality training examples, so the data it learns from gets better and better over time. It’s learning from its own successes.

When tested on a benchmark called MobileWorld (which includes 201 tasks across 20 different apps), MAI-UI achieved a 41.7% success rate. That might not sound like 100%, but it's a massive 20-point jump over the next best models, showing this method really works, especially for tasks that require back-and-forth conversation or using external tools.

Real-Time Training in a Virtual Android Universe

Here’s the part that really got my attention. Learning from a static dataset is one thing, but the real world is messy and apps change all the time. To build a truly robust agent, you need it to practice in a live environment.

So, the team built a massive, scalable "training gym" for MAI-UI. They use containerized Android Virtual Devices—basically, thousands of virtual phones running in the cloud. The AI gets to interact with these virtual phones in real-time, trying tasks, making mistakes, and learning directly from the consequences (a process called Reinforcement Learning).

What they found was that scale matters. A lot.

• When they increased the number of parallel virtual phones from 32 to 512, the AI's success rate jumped by 5.2%.
• When they let the AI take more steps to solve a problem (from 15 to 50), its success rate went up another 4.3%.

It's just like a human learning a new skill. The more you practice, and the more complex scenarios you practice with, the better you get.

This intensive training is what led to the headline-grabbing result. On the AndroidWorld benchmark, which tests live navigation skills, the largest MAI-UI model hit a 76.7% success rate. That officially puts it ahead of UI-Tars-2, Seed1.8, and even Google's Gemini 2.5 Pro.

What This All Means for You and Me

This isn't just an academic achievement. What the Alibaba team has built is a blueprint for the next generation of AI assistants—ones that are genuinely useful.

Here are the big takeaways:

• A Unified Agent for Your Phone: MAI-UI is designed for the real world. It can talk to you, use other tools, and handle your data smartly, not just run in a lab.
• Best-in-Class Performance: The numbers don't lie. It's setting new state-of-the-art results in both understanding what's on your screen and navigating through it.
• Ready for Real-World Messiness: The MobileWorld results show it can handle tasks that require conversation and tool use, which is what most of our real-life phone usage looks like.
• Practice Makes Perfect: The scalable, real-time training environment is a game-changer, proving that giving these AIs more "live" practice is the key to unlocking their potential.

We're still in the early days of AI agents that can truly operate our devices for us. But work like this, focusing on practical challenges like user interaction, tool use, and scalable training, is a massive leap in the right direction. It’s one of the first times I’ve seen a system that feels like it’s being built not just to win a benchmark, but to one day be a genuinely helpful partner on the device we use most.