Qwen Intelligence is Here: Mobile AI Agents

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: Qwen Intelligence is Here: Mobile AI Agents · Channel: Fahd Mirza · 07:37 · Watch on YouTube

Topics: AI agents, AI evaluation and benchmarks, AI models, Local and open-source models

Summary

The video introduces Alibaba’s mobile AI agents: a planner that breaks goals into steps, a UI agent that operates apps, and a creative agent that generates images. The presenter demonstrates the planner creating a plan to set an alarm and send a message, then reviews Alibaba’s benchmark claims and example workflows. He also explains Alibaba’s three-stage development loop and shows the UI agent’s capabilities and demos. He notes that the tools are not open source, the planner demonstration only planned rather than executed the task, and he hopes to test more if open weights become available.

Key points

Alibaba’s system is aimed at phone makers, with Honor described as its first partner.
The three agents are a planner, a UI-use agent, and a creative agent.
The planner demonstration produced steps for setting a 7:00 a.m. alarm and texting the presenter’s wife, but did not execute them.
The presenter says the planner uses the Qwen 3.8 model.
Alibaba’s Mobile PA benchmark chart reports a score of 77.1 for its 27-billion-parameter planner; the presenter notes these are Alibaba’s own figures.
The presenter says the planner repository contains blog and report material, but no released weights or code.
Examples shown include memory-based workflows, sub-agent coordination, personalized state correction, and recovery from errors.
Alibaba describes an AI development loop involving data, training, and hardness, with feedback from real devices.
Alibaba’s charts report scores of 82.1 on Mobile World and 92.2 on Mobile World Real for its UI agent.
The presenter says the showcased scenarios include image creation, autonomous decisions, mini-app development, and long-term task delivery.

Chapters

Start	Chapter	Summary
00:01	Introducing Alibaba’s mobile AI agents	The presenter outlines the planner, UI, and creative agents and explains how to access the planner console.
01:06	Planner demonstration and benchmark claims	He tests the planner with an alarm-and-message task, reviews Alibaba’s benchmark chart, and notes that the repository has no released code or weights.
03:15	Example workflows and development loop	The video covers example agent workflows and Alibaba’s three-stage AI loop for data, training, and hardness.
05:29	UI agent, showcases, and open-source hopes	The presenter reviews the UI agent’s claimed capabilities and demos, summarizes Alibaba’s showcase scenarios, and calls for open weights.

Quotes

This one is built to make your phone actually do things for you. — 00:01

Would be good if they would have open-sourced it right away, but not yet. — 01:06

No weights or code are released, which is a bit of a shame. — 02:44

Alibaba calls it an AI for AI loop with three stages. — 04:53

I tried the planner live, and without tools it planned it planned but it did not execute. — 07:37

Concepts and entities

Qwen Intelligence, Qwen Planner Agent, Qwen UI Agent, Qwen3.8, Mobile PA-bench, Honor, Alibaba, Fahd Mirza

Full transcript (original language)

00:01 Alibaba has launched Guan Intelligence and it is competing right with Apple Intelligence. This one is built to make your phone actually do things for you. It ships with three agents, a planner that breaks a goal into steps, a use agent that operates app, and a creative agent that generates images. It’s aimed at phone makers with Honor as the first partner. Today, I’m not only going to do a bit of hands-on with a planner agent, but also

00:32 unpack all of these because I’m more than sure you are going to see this getting embedded in your phone, especially the Android ones, very, very soon. This is Fahad Mirza and I welcome you to the channel. So, what exactly is this thing? Let me take you straight to their console, which I will drop the link in video’s description. From there, either you can go to their Chinese site and there is also an international site, but the main thing is that you would need to log into their console, create a free account and from there, you can create this planner agent. And

01:06 by the way, this video is not sponsored by Alibaba or anyone. Would be good if they would have open-sourced it right away, but not yet. So, you can see that I have already created a planner agent after clicking on plus. Once your instance is created, then you will be presented with this screen where you can configure your agent. So, you can see just as a test in this planner debug panel, I have given it one-line role, just specifying system instructions, and then a two-part task

01:38 where I’m also going to ask it to set an alarm for 7:00 a.m. and text my wife I’ll be late. I haven’t set any phone number or anything. I’m just clicking here and hopefully it is going to plan something for me. It is using the Guan 3.8 model behind the scene. So, there you go. You can see that it has done all the planning. So, I’ll just quickly show you that this is a clear step-by-step instruction, recipient message text, and these are the steps.

02:08 And then this is a final confirmation. And you can also use any Google Assistant, Siri, or any other voice assistant if you like. And you can see that they have released the Gwen planner agent repo, and this chart is Alibaba’s own benchmark. Mobile PA bench, their 27 billion planner scored 77.1 overall, just behind GPT-6 Astra at 76.8 on the chart scale, and ahead of uh Claude Opus 5 and Claude Fable 5. At a much, much lower output

02:44 cost, which is no surprise at all. But remember, these are Alibaba’s own number, not mine. But one thing to note that this repo is only blog and report site. No weights or code are released, which is a bit of a shame. I think Alibaba could have done better. Otherwise, it is just like Apple Intelligence where everything is a black box. Now, um another thing is that they have shared some of the examples on their page, which is quite interesting. And this is where you see the real

03:15 promise of this tool. So, you can see that this is a memory guide planning. In this one, the planner was told to run its saved dev workflow. The agent searches the memory, recovers procedure, then opens the editor, connects shared input, and then enable USB debugging in order as connected to your your phone. The second one they have shared is also very interesting. A tool error contradicted an earlier reply, and the agent trusted the tool result, stopped the active task, then

03:47 reconnected to the bedroom scene. And then we have sub-agent coordination. This one waits until WhatsApp downloads are confirmed, then hands accept file names to file manager to sort into math, science, and English folder. This one I think uh is quite targeted one. Personalized state correction, where it checks like to speak and talk back, compares them to the user saved routine, and fixes only the mismatch without touching the font. Then they have also got this skill

04:22 guided execution. So, before powering off, it confirms the note is actually saved. Shutdown waits for that confirmation. This might be a bit tricky. And then we have last error recovery after two unit conversion failures, it switches to a calculator, and then still takes the core converts the weight in pounds. So, pretty good. Next up, another thing I wanted to show you is this life cycle, which I believe is the best slide in this whole thing.

04:53 This is how they built it. Alibaba calls it an AI for AI loop with three stages. First, AI for data, agents write executable tasks, collect and verify trajectories, and use failure analysis to decide what data to build next. Second, AI for training, a method called care adjust rewards to the model skill level, balancing finishing the task with doing it efficiently. Third, AI for hardness, skills matched to the available tools and memory that keeps its source evidence, and replaces outdated facts, and then real device

05:29 feedback loops back into all three. Let me now show you this Gwen UI agent, which again comes up with their own blog post. Nothing much is available from their um repo. Again, you need to buy some tokens in order to get it working, which is again not really quite like, but anyway. So, you see that alongside the planner, there is this Coin UI agent, the one that actually operates screen. It works across mobile, desktop, and the web. On

05:59 Alibaba’s own charts, it leads mobile world at 82.1 and mobile world real at 92.2. And this is ahead of the rivals they have picked. Again, there are These are their numbers. So, check the benchmark yourself. And these are all the benchmarks I was talking about. A lot of data has been shared around capabilities. And these are all the demos if you’re interested. I’ll just play one of them. So, this is how it works. Again, mainly

06:32 focused and geared towards Chinese market at the moment, it seems. So, this is how you see it is doing some shopping, browsing through the website. This field is still evolving. So, I think Alibaba is also trying to capture this computer use and browser use category. So, before I wrap up this video, here is what Coin Intelligence says it can do. It has uh four themes, effortless creation where one prompt

07:03 gives you a finished image. Then we have autonomous decisions where it thinks first and acts only when needed. Then they have mini app development where one prompt makes a playable app. And long-term task delivery where it remembers context and keeps a small multi-week goal moving. These are Alibaba’s own showcase scenarios. And um I think it would be good if they provide some open-source open weight repo where we could actually test it. So, that’s Coin Intelligence overview for you, a planner that thinks, a use agent that

07:37 acts, and a creative agent that makes. I tried the planner live, and without tools it planned it planned but it did not execute. If you want me to test the red rest, just wait for it while they open source it and open wait it and we will get them installed and then do a hands-on demo. Recently, I also started this free weekly AI newsletter which you can subscribe at farazmirza.substack.com and if you want to help out the channel, please become a member. Thank you for all the support.