Frontline Hotspot
Frontline Hotspot

Three agents in your pocket: Alibaba's phone AI stack

In September 2026 Alibaba's Qwen team released Qwen Intelligence, a full-stack AI solution for phone makers: it builds no hardware and instead supplies OEMs with models, a platform and scenario-specific agents. The stack contains three agents. Mobile Planner handles planning and decision-making, supporting long-horizon tasks of 100+ steps with a dual-layer memory of working memory plus a long-term user profile. Mobile-Use executes cross-app actions with an API-first, GUI-fallback approach, wrapped in three layers of security and privacy control and requiring user confirmation for key decisions. Mobile Creative targets image creation, turning spoken-language requests into finished images. Vendor-stated figures: 91.8 percent overall task accuracy after landing with Honor; Mobile Planner at about 2.41 dollars per thousand task calls; Mobile-Use at 90 percent end-to-end success, 28.8 percent less completion time, 20 percent lower cost and 59.5 seconds per task; Mobile Creative producing a first image in 3 seconds; and a security score 23 percentage points above leading solutions, with a joint safety benchmark built with Fudan's Baize team. On the platform side it opens four evaluation suites covering planning, execution, real devices and safety, is modular and customizable, and supports independent model deployment with unified Skill governance. The article pins its framing: every performance number is vendor-stated rather than independently retested, the competitor table is sourced from the vendor, and pricing, availability timelines, the full partner list, and model versions or parameter counts are unpublished and are written as following the vendor. It also states plainly that this is a B2B stack rather than a consumer app, with consumers reached only through Honor devices and the Qwen AI platform demo area.

Published September 23, 20267 min read
<!-- qwen-intelligence-hotspot | hotspot | Three agents in your pocket: Alibaba's phone AI stack -->

For a decade the phone assistant has been stuck on one step: you have to call it first. Say the right wake word and it lifts its head; say the wrong one and it pretends not to hear you. Expectations collapsed accordingly: weather, alarm, and one song counted as a pass. The assistant was a search box waiting for you to speak.

Qwen Intelligence targets that step. Alibaba's Tongyi Qianwen team calls it a full-stack AI solution for phones, and the official one-liner is that it moves the phone from passive response to proactive service. Taken apart, that is a redistribution of responsibilities: the phone no longer only runs instructions you issue one at a time, it tries to take on a job you never finished describing.

What matters is not that it taps buttons across apps, but that the job was split into three roles, each with its own cost structure and safety policy, behind a claim worth testing: operating a phone is an engineering delivery problem, not a model-capability problem.

First, what this is and what it is not

Qwen Intelligence comes from Alibaba's Tongyi Qianwen team, and the positioning is blunt: it does not build hardware. It sells phone makers a one-stop AI base layer of models, platform, and scenario solutions. This is B2B. The customer is a handset manufacturer, not you.

That goes first because it is where discussion goes off the rails. This is not an app consumers can download. The verifiable consumer footprint is two things: the Honor partnership, which the vendor says reaches 91.8 percent accuracy on composite tasks, and a demo zone on the Qianwen AI platform, with the main site at qwenintelligence.com. It does not plan to build its own phone or bypass manufacturers, which sets its ceiling and puts the real risk on the commercial side.

On undisclosed items: pricing and billing details, launch regions and timelines, the full partner list, and model versions and parameter counts are unpublished. Every performance figure here is the vendor's official number, not our measurement.

Mobile Planner: the project manager inside your phone

Mobile Planner handles planning and decisions, and you can think of it as the brain. It understands a complex request, breaks it into steps, sequences the available tools, and adjusts on the fly based on feedback.

The official example is long-horizon delivery: planning a job past one hundred steps, such as organizing an event. That number deserves a pause. The hard part of a daily instruction was never "open a certain app," but "get this event off the ground," which contains scheduling, venue, invitations, materials, and budget, and any one snagging forces a revision to the plan. Past a hundred steps, the test is no longer decomposition but whether the system can go back and fix what it already decided, which is the weight carried by dynamic adjustment.

Underpinning this is a two-layer memory system: working memory evolves with the task, while long-term memory accumulates a user profile, which the vendor says means it understands the user better over time. Splitting them is sensible: one is task-scoped and short-lived, the other is about a person and spans tasks.

On cost, the official figure is roughly 2.41 US dollars per thousand task calls, which the vendor says is a fraction of what leading models cost. The number means little alone; its value is that it exposes an assumption: this is built for high call volume. If a phone triggers it a few times a day, cost is not a topic. The vendor did not say which models form the comparison set, so it cannot be checked from outside.

For a sense of what on-device agents look like on the open-source side, compare OmniBot, an open-source on-device agent. The two differ on whether the planning core lives on the device or in the cloud.

Mobile-Use: APIs first, screen automation as the fallback

Mobile-Use is the agent that does things, executing operations across apps. Its approach is described as APIs first, with graphical interface automation as the fallback. If an operation can be done through an official interface, it goes through the interface; if not, it falls back to simulating taps on the screen.

Pure screen automation looks universal: it can tap anything in any app, but it is slow, brittle, and expensive: identifying interface elements burns compute, a redesign breaks the flow, and one stray pop-up halts it. Pure interface calls are fast and stable, but coverage depends on someone else opening their interface, and in practice many apps do not. The trade-off balances coverage against stability: finish the stable path, then use the fallback for the long tail. The vendor says the capability extends to browser scenarios and multi-device coordination.

On safety, the official description is three layers of security and privacy controls, with key decisions executed only after user confirmation. That is necessary once an agent operates across apps: if it can confirm on your behalf, fill forms, and submit orders, it can spend your money, a different risk class from misreading a sentence in a chat.

Every metric here is the vendor's official figure. End-to-end task success rate of 90 percent. Task completion time compressed by 28.8 percent. Cost reduced by 20 percent. A single task completed in 59.5 seconds. On the Honor side, the closure rate is also reported at 90 percent.

Cold water is warranted. Duration is an experience metric, and 59.5 seconds is acceptable for a job crossing three or four apps. Success rate is a trust metric, and 90 percent means one failure in ten runs. For an agent that can pay on your behalf, that 10 percent decides whether you have to watch it finish, and therefore whether the product gets used long term. The vendor published no fallback behavior on failure and no breakdown of the task categories covered.

Mobile Creative: one sentence in, a usable image out

Mobile Creative handles image generation. It turns a casually phrased request into a workflow, then produces a usable image from a single sentence. The official figure is 3 seconds to the first image, which the vendor says puts it in the industry's top tier.

Placed beside the other two, a division of labor appears: Planner and Mobile-Use deliver "the job is done," while Creative delivers "the asset is ready to use," so its headline metric is time to first image. In a creative scenario, tolerance for waiting is far lower; wait too long and the user does it manually.

The table below compares the three agents. The metrics column is entirely official vendor figures.

AgentRoleCore mechanismOfficial metrics
Mobile PlannerPlanning and decisions; the phone's brainRequest understanding, task decomposition, tool sequencing, dynamic adjustment; two-layer memory (working memory plus long-term user profile); supports long-horizon tasks beyond 100 stepsAbout 2.41 USD per thousand task calls
Mobile-UseExecutes phone operations across appsAPIs first with GUI fallback; three layers of security and privacy controls; key decisions require user confirmation90 percent end-to-end task success rate; 28.8 percent shorter completion time; 20 percent lower cost; 59.5 seconds per task
Mobile CreativeImage creation with directly usable outputTurns casual requests into a workflow automatically; produces an image from one sentence3 seconds to first image

Platform, evaluation and safety: what is really being sold is a base layer

Looking only at the three agents understates the offering. The official platform capabilities include modular customization, independent model deployment, unified Skill governance, one-stop operations, and automated evaluation. Translated: a manufacturer can buy only part of it, run models on its own infrastructure, and skip building evaluation and operations from scratch.

The most notable item is the release of four evaluation suites covering planning, execution, real-device testing, and safety. Open evaluation sets lower the cost of selection and acceptance, because manufacturers can run the tests themselves instead of taking the vendor's word. That is a substantive step for this kind of B2B offer.

On safety, the vendor says its three-layer controls put its safety score 23 percentage points above leading alternatives, again an official figure, and mentions co-building safety evaluation with Fudan University's Baize team. Doing evaluation with an academic security team gives trust an external anchor, and the direction is right, but the 23-point gap still lacks a reviewable methodology.

Below is the vendor's competitive comparison, taken from the source page repeating vendor claims and not our independent retest. Treat it as a vendor capability statement, not a neutral conclusion.

DimensionQwen IntelligenceApple Intelligence (Siri AI)
Product positioningA full-stack AI solution sold to phone makers (B2B base layer)Built and used in-house by Apple, limited to its own device ecosystem (closed B2C loop)
ArchitectureThree agents split across planning, operation, and creation; models and harness evolve togetherIn-house on-device models plus Google Gemini, deeply integrated into the system
Task executionAPIs first with GUI fallback; 90 percent end-to-end success; 59.5 seconds per taskSupports complex cross-app tasks, but Siri AI has slipped repeatedly and only ships in beta in autumn 2026
MemoryTwo layers: working memory plus a long-term user profilePersonal context understanding (mail, photos, messages), dependent on iCloud sync
CreationCreative agent produces an image from one sentence, 3 seconds to first image, top tier per vendorImage Playground generates images and writing tools are supported, but capabilities are more basic
CostAbout 2.41 USD per thousand tasks, a fraction of leading modelsNot billed separately, but limited to expensive recent models (iPhone 15 Pro and above)
OpennessFour evaluation suites released, modular and customizable, already deployed with Honor (91.8 percent accuracy)Closed ecosystem, developer APIs only, Siri AI currently unavailable in mainland China

One thing must be said when reading this table: it is not the same layer of the stack as Apple Intelligence. One is an AI base layer sold to phone manufacturers; the other is a closed capability built by Apple for its own ecosystem. Comparing them head to head is like comparing a building materials supplier with a property developer. The detailed comparison belongs in this batch's five-way review of mobile GUI agents. This article sticks to the facts.

Our take: what the three-agent split solves, and what it does not

What it gets right. It reframes phone operation from a model-capability problem into an engineering delivery problem. For two years, discussion of on-device agents has clustered around whether the model is smart enough; the answer here is that smartness is not the point. Separating planning from execution, putting interfaces first with screen automation as a fallback, and layering memory are what make an architecture deliverable to a manufacturer. It also puts cost on the table, because a plan that does not discuss cost cannot get approved on a device with daily active users. And open evaluation suites, external security review, and user confirmation on key decisions attack one problem: making a manufacturer willing to hand over system permissions. With cross-app operation, technology was never the first obstacle. Trust was.

What it does not solve. First, B2B means the rollout pace is not Alibaba's to set. It depends on how willing manufacturers are to adapt and how far they open system permissions. The fallback depends on accessibility services, screen capture, and simulated taps, and how much of that opens is decided by manufacturers and platform owners. Without cooperation, it is a capability on paper.

Second, there is too little verifiable public information. The 91.8 percent figure is composite task accuracy, with no detail on how it is calculated or which scenarios it covers. Every number is official, with no third-party retest, so any conclusion today about who beats whom is a vendor statement.

Third, long-horizon reliability is not addressed. The vendor emphasizes planning beyond one hundred steps but publishes no success rate, and says nothing about mid-task failure or rollback. If one wrong step in a hundred sends the job back to the start, the experience and cost math has to be redone.

Fourth, the commercial picture is almost blank. Pricing, billing details, launch regions, timelines, and the full partner list are unpublished. Honor is a starting point, but one customer is not a market.

The conclusion. The value here is a clearly shaped answer: the agent on a phone should be a set of system components with a clean division of labor, countable costs, and controllable permissions, not a smarter assistant. Whether it lives on phones depends on two things it does not control: whether manufacturers open permissions, and whether the official numbers can be reproduced by third parties. Until then, treat it as a direction, not a conclusion.

If you want to feel what an agent on a phone is like to use, the hands-on SOP for putting an agent on Android is a reasonable start.

Sources

  • AI tool directory listing (original source credited to the Qianwen large model), collected on 2026-09-23.
  • All performance figures in this article (91.8 percent, about 2.41 USD per thousand calls, 90 percent, 28.8 percent, 20 percent, 59.5 seconds, 3 seconds to first image, 23 percentage points on safety score) are official vendor claims, not our measurements. The competitive comparison table reflects the source party's claims, not our independent retest.
  • Pricing and billing details, launch regions and timelines, the full partner list, and model versions and parameter counts are unpublished. Defer to official announcements.

This article is AI-assisted and human-edited. Last updated: 2026-09-23

Related

Frontline Hotspot

From 2.8s to 2.3s: can Qwen3.8 steal the interpreter's job?

In September 2026 Alibaba's Qwen team released Qwen3.8-LiveTranslate, a real-time simultaneous interpretation model opened through the Qwen AI platform and Alibaba Cloud Bailian as a WebSocket streaming API that can be embedded in meeting systems, live streams and support desks. Headline figures: average lag (LAAL) cut from 2.8 to 2.3 seconds; recognition input in 60 languages and speech output in 29; three capabilities, real-time speaker diarization plus voice cloning, source and translation emitted in the same frame, and long-context disambiguation, with video and audio input helping resolve ambiguity. Technically it rests on an Interleave single-stream architecture that caches already-heard audio and already-emitted translation instead of reprocessing each sentence, plus a Hybrid MoE Thinker-Talker pair, where the Thinker arranges video, audio, source and translation into one causal sequence and the Talker fuses translation with source audio into speech that keeps the original speaker's timbre. The article keeps its figures honest: 2.3 seconds is average lag rather than end-to-end first-packet latency, 60 and 29 are different units, the vendor comparison table is not independently retested, an unpublished metric is not the same as a bad one, pricing, rate limits, concurrency and regional availability are not invented, and the model is an API service rather than open source.

Sep 21, 20267 min read
Frontline Hotspot

Qwen3.8-Omni-Flash: From Understanding to Delivery, Omni Agent

Per ai-bot.cn on 2026-09-18, Alibaba Qwen released the next-generation natively omni-modal Qwen3.8-Omni-Flash: text, image, audio and video input with a 1M context window, built for "from understanding to delivery" agent workflows such as meeting minutes, drama dubbing translation and movie recaps. Against Gemini 3.8 Flash the picture is honestly three wins and three losses: big wins on AliMeeting DER (3.4 vs 72.6) and WildClawBench-MM (71.0 vs 58.9), while trailing on OmniVideoBench, LVOmniBench and FLEURS WER. API audio input price drops over 98% and audio-video over 93%, pushing audio costs to the floor. The companion Qwen-MM-Plugins is open-sourced, evolving model and toolchain together.

Sep 19, 20267 min read
Frontline Hotspot

Qwen3.8-Flash Free on Qoder: How to Use the 13-Day Window

Per an IT之家 report on 2026-09-18, Qoder launched a double promo: from 10:00 on September 18 through 23:59:59 on September 30, the Qwen3.8-Flash billing coefficient drops from 0.1 to 0, making calls completely free; meanwhile 100 Credits are granted daily from 10:00, valid 30 days and stackable, open to all individual users on both the international and China editions. This piece unpacks why free windows are the standard go-to-market play for coding tools (acquisition, habit formation, post-window retention), maps where Qwen3.8-Flash fits, and reminds readers this is a limited window, not a permanent free tier: the coefficient returns to 0.1 after September 30.

Sep 18, 20267 min read