Yesterday, the official Qwen-UI-Agent platform unveiled a demonstration video.

In the video, the AI first navigates to Douyin (the Chinese version of TikTok), searches for the most bookmarked image and text post about "beef in passion fruit sour soup," bookmarks it, and jots down the ingredients. Subsequently, it transitions to Hema (a popular grocery delivery app), adds all the ingredients—excluding seasonings—to the shopping cart, and schedules delivery for 6:45 PM the same day.

Previously, Doubao Phone had showcased a similarly impressive demo. The user simply speaks a command, and the phone autonomously opens various apps, conducts searches, compares options, and places orders. The key difference is that Doubao integrated this functionality into a 3,499 RMB engineering prototype, only to face restrictions from WeChat, banking, and payment apps due to risk management concerns. Qwen, on the other hand, has documented these actions in research papers and project repositories.

When it comes to cross-app operations, Doubao encountered product risk control issues, whereas Qwen highlighted them in model demonstrations.

Figure 1 | The first frame from Qwen-UI-Agent's official demo.

The instructions on the right instruct the agent to first locate a highly bookmarked recipe on Douyin and then purchase the ingredients on Hema, with delivery scheduled for 6:45 PM. This demonstrates Alibaba's capability to execute cross-app long tasks, though it does not confirm that this functionality has been integrated into mass-produced phones. Source: Qwen-UI-Agent's official project page.

Additionally, Qwen-UI-Agent has shared technical reports, project pages, and demo videos. However, as of today, the weights of its flagship model remain proprietary. The truly open-sourced model is its predecessor, MAI-UI, which has released 2B and 8B versions under the Apache 2.0 license; the larger 32B and 235B versions are not publicly available.

Thus, this is not a strict open-source release. What Alibaba has done is reframe the system-level agent capability—previously embedded by Doubao into an engineering prototype—into a standalone model capability that is trainable, evaluable, adaptable to different devices, and not inherently tied to system-level UI.

Here, open source serves more as a foundation for Qwen's strategic roadmap. The previous MAI-UI model lowers the barriers to experimentation, while the latest Qwen-UI-Agent showcases the upper limits of its capabilities. This move has also, to some extent, influenced the current landscape of AI phone development.

When AI-driven phone control can be embedded into small open-source models, is there still a need for entirely new AI phones?

The background of Doubao Phone requires no further elaboration—its collaborative launch and subsequent bans by multiple apps are well-known.

ByteDance's ambitions extend beyond a mere voice entry point. It aims to position itself as an intermediary between user needs and all apps.

The significance of this move is that Doubao, for the first time, brought system-level agents to real users and clashed with the boundaries of super apps across the industry.

On December 5, Doubao Phone Assistant issued another statement, restricting score-brushing and incentive-brushing scenarios, temporarily disabling operations for banking and internet payment apps, and limiting access to certain competitive games. User authorization does not automatically absolve platforms of liability. If AI clicks the wrong button, reads a verification code, or turns a misclick into a payment, post, or account risk, liability cannot be dismissed with a simple "the user agreed."

Super apps have their own vested business interests.

A GUI agent can bypass homepages, ads, and recommended sections to directly complete searches and transactions. For users, this means fewer taps; for apps, it could mean losing traffic entry points, ad inventory, and transaction allocation rights.

Recently, Doubao has shifted its strategy. LatePost reported that the new Doubao Phone, when interacting with top apps, will no longer read screens or simulate clicks. It can only integrate if the partner actively provides MCP services, opening relevant data and operation permissions.

MCP can be understood as a toolkit provided by the app itself, specifying which data can be read, which actions are permitted, and which steps require user confirmation. Agents do not need to access excessive underlying data, and users do not have to worry about AI over-operations.

The new Doubao Phone reportedly increased its stock from around 30,000 units in the previous generation to hundreds of thousands, with ByteDance still in negotiations with phone manufacturers and top apps. In December 2025, Jiemian News reported ByteDance advancing AI phone collaborations with hardware makers like Vivo, Lenovo, and Transsion; Lenovo later confirmed ongoing close communication.

At this point, ByteDance's key negotiating leverage has shifted from "I have a large model" to "I've built a phone that truly operates across apps." It has navigated system permissions, real-device adaptation, and app risk controls.

But it can no longer claim exclusivity.

Alibaba Disassembles a Phone into Models

Less than a month after Doubao Phone encountered risk control issues, on December 29, 2025, Alibaba's Tongyi Lab released MAI-UI and open-sourced the 2B and 8B model weights.

What Qwen does is quite similar to Doubao: the model receives user instructions and the current screen, decides whether to click, swipe, input, or call an interface; after execution, it checks the new page to correct any task drift.

The difference is that Doubao encapsulated this loop into a product relying on system permissions. Alibaba, however, disassembled the loop into downloadable, trainable, and deployable models.

In short, Alibaba open-sourced GUI phone capabilities.

The newly released Qwen-UI-Agent takes this a step further. The paper states that the team built an environment with over 100 real devices and 150+ apps, generated around 10,000 simulated environments in parallel for training trajectories, and used online reinforcement learning to handle long tasks exceeding 100 steps. Its actions are not limited to screen taps: for files, spreadsheets, and data processing, the model switches to command-line; multiple sequential operations can be generated in one output.

Figure 2 | Paper Figure 2.

A flight cancellation notice triggers the agent to find alternative itineraries, then, with user consent, modify meetings and spreadsheets on a computer. The left side lists phones, computers, browsers, and action spaces like GUI, command-line, and API. Source: Qwen-UI-Agent Technical Report, p.5.

The paper's value lies in expanding "phone tapping apps for users" into a cross-device execution system. Tap screens when needed, use interfaces when appropriate, and resort to command-line for file processing. GUI is no longer the only path, nor the sole decision-making route.

Official benchmarks look impressive: Qwen-UI-Agent achieved 82.1%, 92.2%, and 97.5% success rates on MobileWorld, MobileWorld-Real, and AndroidDaily, respectively, and 79.5% on OSWorld-Verified. The previous MAI-UI-235B scored 76.7% on AndroidWorld, outperforming UI-TARS-2's 73.3%.

But benchmarks alone do not tell the full story.

The latest mobile scores primarily come from Alibaba's own built or operated evaluation environments. Some competitor scores in the table are also rerun by the team, not equivalent to public test submissions from all companies. By Qwen-UI-Agent's generation, the official-promoted mobile leaderboard shifted from AndroidWorld to MobileWorld, MobileWorld-Real, and AndroidDaily, which better reflect Alibaba's focus on real-device and long-task capabilities but make external verification harder.

The paper itself does not overstate claims.

It documents a task where the agent must add sub-10 RMB items in a Popo Supermarket livestream to the cart, then bookmark the priciest item. The model repeatedly clicks bread, durian, and matcha cake but fails to comply. Another task requires the agent to post a WeChat Moment, then restrict visibility to "only myself"; the post succeeds, but the privacy setting fails.

Figure 3 | Paper Figure 23.

The top half shows failed item selection in Popo Supermarket; the bottom half shows a posted Moment with a failed "only myself" privacy setting. Both cases are labeled MODEL FAILURE by the team. Source: Qwen-UI-Agent Technical Report, p.53.

Compared to benchmarks alone, these failures feel more authentic. Longer tasks amplify early misjudgments into complete failures; posting, paying, and modifying permissions are precisely where users least tolerate AI errors.

Qwen proves Alibaba has built a machine to collect failures, train models, and validate on real devices. It hasn't yet proven a Qwen-equipped mass-produced phone can stably manage users' lives, but this attempt is far from insignificant.

Open source must also be dissected.

MAI-UI only released "small" 2B and 8B versions; the 235B version that scored 76.7% on AndroidWorld remains closed. The latest Qwen-UI-Agent's flagship weights also lack download access. Alibaba Cloud has integrated Qwen models and cloud-based Android containers into its Agentic Mobile product for Wuying Cloud Phones, charging by instance specs and usage duration. This isn't the same product version as Qwen-UI-Agent but clearly outlines a commercial path.

Open-source "small cups" drive traffic; cloud-based "large cups" monetize.

In March, when Qwen lead Lin Junyang departed, the main concern was whether Qwen's open-source roadmap would end. Wu Yongming later confirmed in an internal letter that Alibaba would persist with open source, with Zhou Jingren continuing to lead Tongyi Lab. Zhou's name also appears in MAI-UI's technical report author list.

Though personnel changed and usage evolved, the open-source roadmap remains.

Previously, Qwen open-sourced primarily to gather developers around its models. For phone agents, open source serves a more specific purpose: when negotiating exclusive collaborations with phone makers, Alibaba can present a testable, modifiable, and comparable version.

It doesn't need to outpace rivals immediately. Phone manufacturers now have a second technical reference.

This is a negotiating tool.

For phone makers, MAI-UI's 2B and 8B weights are far from a Doubao Phone equivalent.

Beyond models, they must address system permissions, real-device data, app adaptation, security policies, account systems, and post-sale liability. An AI misclicking a menu in a demo is a model failure; a misclick during payment across millions of devices means customer service calls for the manufacturer.

But user choices have evolved.

Previously, a manufacturer unwilling to train a GUI agent from scratch could partner with ByteDance to integrate Doubao. Now, they can download MAI-UI for prototyping, accumulate data, adapt systems, and then decide which stronger capability to procure. They may not use Alibaba's open-source model but gain a technical reference for comparing prices and terms.

Alibaba seeks to dilute exclusivity.

On July 13, Step AOS and prototype STEPX Neo were released by Jueyue Xingchen, with Alipay, Meituan, Ctrip, and JD as initial ecosystem partners. Different companies use different names, but their actions increasingly align: model firms aim to bypass individual apps and become cross-brand, cross-service intermediaries.

Phone makers hold system permissions, super apps hold transactions, social graphs, and data, and model firms hold understanding and execution capabilities. All want agents to follow their lead but refuse to cover others' mistakes. The likely outcome is a hybrid approach: long-tail apps let AI view screens, high-frequency and sensitive actions use MCP or A2A, and payments, posts, and permission changes return control to humans.

ByteDance won't retreat. With experience co-developing systems for ZTE and user-product closures from Doubao and Douyin, no public evidence yet shows how Honor, OPPO, Vivo, or Xiaomi will choose. But clearly, the AI upstream environment for phone makers has subtly shifted.

In the coming months, if Qwen releases flagship weights or a mainstream phone maker adopts MAI-UI publicly, Alibaba's open-source roadmap will move from GitHub into the phone business. Conversely, if Doubao secures structured interfaces from WeChat, Taobao, Alipay, and other top apps, turning hundreds of thousands of stock units into stable products, ByteDance's early tuition could become first-mover experience.