China's AI Assistant Apps | AI Wars of 2026
Consumer-facing assistants from Chinese AI providers. Source-backed analysis from AI Wars of 2026.
China's AI competition extends beyond models to consumer assistants.
The consumer-assistant contest concerns what people can do with a model: find information, understand a document, create something, or hand over a piece of work. Qianwen, Wenxin, Doubao, DeepSeek, Kimi and Yuanbao offer different combinations of these activities. Their product descriptions show an expanding range of tasks, but they do not establish which assistant completes those tasks most reliably. Source 1 Source 2 Source 3 Source 4 Source 5 Source 6
The strategic question is whether an assistant becomes a service people choose repeatedly because it helps them finish useful work. A model release can improve the assistant's underlying capabilities. It does not, by itself, show that the surrounding product makes those capabilities accessible, controllable, or worth returning to.
The app is the unit of comparison
This chapter examines consumer-facing assistants from Chinese AI providers. It is not a census of AI use in China, a model leaderboard, or a claim that every listed feature is available to every user. A provider, a model family, and an assistant app are different units. Mobile, web, and desktop versions also need to be distinguished when their descriptions differ.
Tencent provides a useful example: its Hy4 announcement names Yuanbao, WorkBuddy, CodeBuddy, ima, and API services as distribution routes. The same model appearing in several products does not make those products one assistant or give them the same access to user information. Source 6
The distinction matters in the other direction too. Kimi's help center describes consumer chat, a general agent, a coordinated agent mode, and separate coding and local-work tools. A capability documented for one mode should not be assigned to every conversation carrying the Kimi name. Source 5
What the products offer
The descriptions below are the providers' stated capabilities, not results from a shared performance test. App-store listings identify particular clients and storefronts; they do not establish availability for every region or account.
| Assistant | Documented product emphasis | Distinction that matters |
|---|---|---|
| Qianwen | Search, webpage summaries, presentations, image generation, and recording notes. | A broad assistant feature set does not establish transaction access or integration with every Alibaba service. |
| Wenxin | Chinese-language assistance, multimodal inputs, writing, presentations, and multi-step task delegation. | Producing a deliverable is a different promise from answering a question. |
| Doubao | Mobile voice and photo interaction, creative tasks, and work outputs; a separate Windows offering describes desktop and browser operation. | The desktop action claims should not be treated as mobile-app capabilities. |
| DeepSeek | A consumer conversational assistant for search, writing, reading, problem-solving, and translation, with Apple and Android listings. | The consumer app needs its own evidence; model or API results do not establish its task outcomes. |
| Kimi | Search and reasoning in chat, plus agents for research, coding, slides, and document or spreadsheet work. | Chat, general-agent, coordinated-agent, and developer tools have different scopes. |
| Yuanbao | A consumer-assistant route through which Tencent distributes Hy4 preview. | Access to a model does not establish shared context or permissions across Tencent's other products. |
Sources for the product distinctions: Source 1 Source 2 Source 3 Source 7 Source 4 Source 8 Source 5 Source 6
Qianwen: everyday assistance and content creation
Alibaba's official Qianwen page describes AI search, webpage summaries, presentations, image generation, and recording notes. That puts information gathering and content production within the assistant's stated scope. It does not establish that the assistant can complete a purchase, change a booking, or act across Alibaba's wider services. Source 1
The strategic opportunity is continuity between those tasks: understanding material and then producing something useful from it. The test is whether a user can carry the relevant facts into the next step without losing their meaning or checking everything again. A list of separate features cannot answer that question.
Wenxin: from answering to delivering work
Baidu's China App Store listing describes text, voice, image, video, and document inputs. Its task-delegation feature is presented as planning and executing multi-step work, including research reports, presentations, data organization, and visualizations. These are Baidu's descriptions of the product, not independently measured completion results. Source 2
That promise changes the relevant test. For a report, fluent prose is not enough: the evidence must support the claims, the required sections must be present, and the result must be usable. For a presentation, a complete file can still fail if its figures or citations are wrong.
The listing also describes uploading documents from WeChat chat records. That is a specific file-transfer route. It should not be read as unrestricted access to a user's conversations or as proof that Wenxin can act throughout WeChat. Source 2
Doubao: distinguish the mobile assistant from desktop action
Doubao's Apple listing advertises information lookup, trip planning, study assistance, data analysis, code writing, and document or slide generation. It also describes voice input, voice calls, recognition from photographs, and image or video creation. Its own description warns that outputs may misunderstand or mislead and should be checked against other sources. Source 3
The Windows listing makes a broader action claim: local-computer operation, browser use, skills, and scheduled tasks. It also describes continuing conversations across devices. Those statements concern the desktop product; they do not establish that the mobile app has the same execution access. Source 7
This makes Doubao a useful example of why app comparisons need a named surface. Asking a question by voice and authorizing changes on a computer are different activities. The latter requires evidence about permissions, confirmation, failure handling, and recovery, not merely a longer feature list.
In September, Doubao added two developments to the mobile story. It released a consumer mobile AI assistant on the Nubia NaviX Ultra phone with the Screen Automation Execution Protocol: third-party apps opt in to permit AI-driven actions, which makes the permission model an explicit product choice rather than a hidden background behavior. Doubao also set an independent 12% transaction fee on in-assistant hotel orders, 11.4% service plus 0.6% payment, compared with 8% on Douyin's main site. This is the first independent pricing of AI-conversation traffic in China, and it values the assistant channel as distinct from the parent platform's traffic. Both are company-reported developments without independent verification. Source 9 Source 10
DeepSeek: evaluate the assistant separately from the model
DeepSeek's Apple listing identifies an official consumer assistant and advertises search, writing, reading, problem-solving, and translation. Google Play separately identifies its official Android assistant. These records establish product identity and advertised tasks, not a comparative measure of answer quality or successful work. Source 4 Source 8
The company's own release log shows how quickly the underlying model changes. It lists general availability of its V4-Pro model in the app, web, and API in August 2026 and a smaller V4.1-Flash model in September. These are company statements, and the benchmark scores in the log are the company's own. A model release changes what the app can do; it does not show how well the app serves a person's task. Source 11
For this product, the important distinction is between choosing a model and choosing the application around it. A useful app assessment asks whether answers can be checked, whether the conversation helps the user reach a sound result, and what work remains for the user. Model research cannot substitute for those observations.
Kimi: research and production across distinct modes
Kimi's help center describes built-in web search, reasoning, and file handling. It distinguishes chat from a general agent for research, website generation, slides, documents, and spreadsheets. It also describes coordinated agents for larger searches and batch processing, while presenting its coding and local-work products separately. Source 5
Kimi's site also announces its K3 model, described as built for agentic coding and knowledge work. That is a product claim, not a test result. Source 12
Kimi's official pricing page lists four monthly subscription tiers from ¥49 to ¥699, with lower monthly prices when billed annually. Every tier draws from one shared pool of usage credits across the product's modes, including K3 and Kimi Claw. Kimi Code also has a separate weekly limit. Hosting a Kimi Claw sandbox or a published Agent website consumes credits each day. These terms describe how Moonshot allocates paid usage. They do not establish value, retention, or task results. Source 13
Kimi therefore belongs in more than one operating pattern. The relevant comparison depends on whether the user wants an answer, a researched deliverable, or work that includes executing actions. More agents are not automatically better: any additional work still has to preserve source support and produce a result the user can inspect. Assistant Architectures explains these distinctions in more detail.
Yuanbao: a model-distribution route, not a unified permission system
Tencent's Hy4 announcement explicitly includes Yuanbao alongside its productivity, coding, knowledge, cloud, and API surfaces. This documents a route for bringing a model to users through several products. The announcement's model-level evaluations do not establish Yuanbao-specific reliability, repeated use, or customer outcomes. Source 6
The strategic possibility is that users can reach useful AI within a familiar set of services. The unresolved question is what travels between them: a conversation, a selected document, a result, or permission to act. Model availability alone answers none of those questions. The broader Tencent chapter examines that company-level route; Yuanbao still needs product-specific evidence.
Office and desktop agents are being consolidated
Agent products aimed at office work are changing faster than the consumer apps around them. QuestMobile's September report describes Tencent, Alibaba, and ByteDance each bringing several office agents into one group-level offering. Source 14
Tencent's closure of QClaw shows what that can mean for a user. Its notice says QClaw will stop operating on 24 December 2026, and that sign-ups and purchases stopped on 24 September. Configuration, memory, agent persona, and conversation history can move to WorkBuddy — but only to a personal WorkBuddy account. Source 15 In this case, what moves between the two products is set by the company's notice and limited to one account type. A comparison of agent products therefore needs a date, and should record whether each product still exists.
ByteDance's TRAE brand adds another office assistant to this comparison. Its official page describes TraeWork as an AI work assistant for presentations, data analysis, research, document writing, and code development. Work, Code, and Design modes synchronize across desktop and web. This is ByteDance's description of the product, not evidence of repeated use or results. Source 16
Baidu added a fourth consolidation event at its September 21 AI Day: its overseas AI office product, formerly Oreate AI, was rebranded as Kooko. The company claims 10 million overseas users in the first year across roughly 200 countries, with paying customers in 197 countries and monthly recurring revenue growing 50% month over month. The product's domestic counterpart, 库库AI, reports more than 40 million monthly active users. These are company claims reported through technology press, not independently verified measures. Source 17
Zhipu's AutoClaw is a separate product in the same category. Its official page lists tools for office documents, social-media content, financial analysis, website and mini-program creation without code, browser automation, and messaging integration. This describes product scope. It does not supply an activity figure comparable with WorkBuddy, QClaw, or TraeWork. Source 18
Distribution helps people arrive; useful work gives them a reason to return
The working hypothesis is that an assistant becomes harder to replace when people can repeatedly move from a question to a useful, checked result. That could happen through a broad set of everyday functions, a strong research-and-writing workflow, or a well-controlled action mode. The product records describe candidates for these routes, not established winners.
Distribution and integration must also be kept separate. A download link shows a way to obtain a client. A file-import feature describes a way to bring information into it. A shared model connects products technically. None of these alone establishes a continuous workflow with the user's permission preserved at each step.
This is why the consumer-app comparison complements rather than replaces the Baidu, Tencent, and ByteDance company chapters. A company's broader reach is context for an assistant's opportunity. It is not an app-level outcome.
Permissions and data use are part of the product choice
An assistant's usefulness depends partly on what information the user is willing and able to provide. Uploading a document, sharing a photograph, continuing a conversation on another device, and allowing desktop actions create different questions about access and control.
DeepSeek's Google Play Data safety section contains developer declarations about data collection, sharing, encryption in transit, and the ability to request deletion. Those statements are useful disclosures, not an independent audit of storage, retention, or deletion effectiveness. Source 8
Doubao's Windows listing illustrates another limit: a description of computer-operation capabilities does not explain every permission prompt, approval requirement, or recovery mechanism. Neither listing supports a comparative privacy or safety ranking. Source 7
China's rules for anthropomorphic AI interaction services, in force since 15 July 2026, have a narrow scope. They apply to services that provide sustained emotional interaction by simulating a person's personality and manner. Customer service, question answering, work assistants, and education and research services are outside them unless they involve that kind of interaction. Source 19 Source 20 Most of the tasks compared in this chapter fall outside that scope, but an assistant offered as an emotional companion could fall inside it.
For a specific task, the practical questions are narrower: what information must be supplied, what the assistant may read or change, what requires confirmation, and what can be recovered if the result is wrong.
What the app-activity figures show
QuestMobile's half-year report on AI applications is the original source of the activity figures that news sites repeat. It dates its category total to June 2026. Two news summaries date the same total to May, so the report, not the summaries, is the reference for the date. Source 21 Source 22 Source 23
The report measures monthly active users, uses per person, and minutes per person for named apps. Its listed apps together exceed its category total. That is consistent with people using more than one app, but the report does not say how that overlap is counted. A second data provider ranks the same apps for the same month but reports different sizes. The differences vary by app, and its product set differs too. Source 24 Neither provider states a method that would allow the two to be reconciled. No activity figure or relative ranking is used here.
An earlier QuestMobile report provides a separate snapshot for March 2026. It reports 440 million monthly active users for the AI-native-app category, with 345 million for Doubao, 166 million for Qianwen, and 127 million for DeepSeek. It also gives quarterly average “activity rates” of 33.5%, 17.1%, and 21% for the three apps. But it does not define that measure or relate it to monthly active users. Only the report's text summary was available; its charts were not read. These figures are not used as market shares or a ranking. Source 25
The distinction is not only statistical. An active app user, time spent in a session, repeat use by the same person, and successful completion of a task answer different questions. More time in an assistant could reflect useful work or difficulty getting the desired result. An activity estimate cannot decide between those explanations.
The same report shows why a user count is a weak guide to use. Two of the largest general assistants have user counts of a similar size, yet time per person differs several times over. More users therefore do not show that an assistant is used more. Source 21
TechNode and IT Home relay the same underlying QuestMobile study; they are not independent measurements of the market. Until the providers' counting rules can be compared, the estimates should not determine the chapter's competitive assessment. Source 22 Source 23
Compare the apps using work a person actually needs done
A useful comparison would start with the same task, inputs, and permission limits, rather than the longest list of advertised features. Examples include:
- Research: answer a question from a defined source set, checking whether each citation supports the claim attached to it.
- Documents and presentations: turn supplied material into an editable deliverable while preserving figures, qualifications, and attribution.
- Everyday assistance: explain supplied material clearly, noting where a wrong answer would require expert or independent checking.
- Desktop action: complete an authorized change within a named environment, recording human interventions and whether errors can be undone.
These are proposed tests, not results. Useful measures would include completion quality, unsupported claims, correction effort, permission failures, and whether people return for the same type of work. The app, mode, platform, task, and observation period must be named.
Decision question
Which assistant can turn its documented product capabilities into repeated, reliable, user-authorized work, and what evidence would distinguish that achievement from app availability or a provider's description?
What would strengthen the case
Watch for product-specific evidence that connects the route to a result. That means repeat use for a defined task, independently checked deliverables, reliable handoffs between stages, and effective controls where the assistant can act. A published rule for counting people who use several apps, applied by both providers, would strengthen activity comparisons. It would not by itself establish task performance.
The chapter's current conclusion is bounded but useful: these assistants offer different ways to begin and organize work, so app and mode selection matter. The evidence does not yet establish a general performance leader or show which route will become a durable advantage.
Sources
- Source 1
- Source 2
- Source 3
- Source 7
- Source 4
- Source 8
- Source 12
- Source 5
- Source 6
- Source 22
- Source 23
- Source 21
- Source 14
- Source 25
- Source 24
- Source 11
- Source 15
- Source 13
- Source 16
- Source 9
- Source 10
- Source 17
- Source 18
- Source 19
- Source 20
Source references
Source 1: Qianwen - Alibaba AI Assistant
Alibaba
Qianwen's official page identifies it as Alibaba's AI assistant and lists AI search, webpage summaries, AI presentations, image generation, and recording notes. This is product identity and company-described functionality, not evidence of adoption, repeated use, reliable task completion, or user outcomes.
Source 2: Wenxin - Baidu's AI Assistant App
Baidu
Baidu's China App Store listing identifies Wenxin as its AI assistant and describes a task-delegation feature that plans and carries out multi-step work. This is a company product claim, not independent task-performance, reliability, adoption, or outcome evidence.
Source 3: Doubao - AI Assistant: Apple App Store Listing
Beijing Chuntian Zhiyun Technology Co., Ltd.
The named publisher's Hong Kong App Store listing advertises information lookup, trip planning, study assistance, data analysis, code writing, documents and slides, voice input and calls, photo recognition, and image/video creation. It warns that outputs may misunderstand or mislead and recommends checking other sources. Inspected September 28, 2026; a relative release date is not converted into a publication date. The listing is publisher-authored product information, not tested capability, outcome evidence, nationwide availability, or proof of a formal corporate ownership relationship.
Source 4: DeepSeek - AI Assistant: Apple App Store Listing
Hangzhou DeepSeek Artificial Intelligence Co., Ltd.
The publisher's Macau App Store listing identifies DeepSeek's official consumer assistant for iPhone and iPad and advertises search, writing, reading, problem-solving, and translation. The page was inspected on September 28, 2026; its current release date was relative, so no absolute publication date is assigned. Product identity and advertised tasks do not establish tested functionality, comparative performance, retention, or availability for every country or account. Historical image/file-upload release notes were not tied to a verified current version and are not used as current capability evidence.
Source 5: Kimi Overview
Moonshot AI
Kimi's undated official help overview, retrieved September 28, 2026, distinguishes chat and model selection from a general agent for research, website generation, slides, documents and spreadsheets, coordinated agents, Kimi Code, and local-work tools. It describes web search, reasoning, file handling, browser access, and mobile clients. These are provider descriptions of different product modes, not a claim of feature parity, measured task success, superiority from additional agents, or availability to every account. Numeric capability and performance claims on the page are not promoted into the chapter.
Source 6: Tencent Releases and Open-Sources Tencent Hy4 Preview
Tencent
Published: 2026-08-28
Tencent announces Hy4 preview availability through WorkBuddy, CodeBuddy, Yuanbao, ima, Tencent Cloud TokenHub, and OpenRouter. The announcement also reports an internal evaluation. It establishes distribution routes and stated product availability, not independent reliability, repeated use, or commercial outcomes.
Source 7: Doubao - Windows App: Microsoft Store Listing
Beijing Chuntian Zhiyun Technology Co., Ltd.
The publisher's Windows Store listing describes conversation, web search, research reports, creative generation, and presentation/document/spreadsheet handling. It separately advertises local-computer operation, browser use, skills, scheduled tasks, and cross-device conversation-history synchronization. The displayed last-updated date is August 19, 2026. These desktop-specific claims are not assigned to the mobile app. A broad Store capability label does not establish the precise permission prompts, approval requirements, sandbox boundaries, recovery mechanisms, reliability, or actual access behavior.
Source 8: DeepSeek - AI Assistant: Google Play Listing
DeepSeek
The publisher's Google Play listing identifies the official Android assistant under package com.deepseek.chat and shows an update date of September 17, 2026. Its developer-supplied Data safety declarations describe possible data collection and sharing, encryption in transit, and the ability to request deletion; the listing states practices may vary by use, region, and age. These are declarations, not an independent audit of storage, retention, training use, encryption at rest, or deletion effectiveness. The listing does not establish product performance or universal availability.
Source 9: Doubao launches consumer mobile AI assistant with Screen Automation Execution Protocol
AI Market Watch
Published: 2026-09-16
Reports Doubao's consumer mobile AI assistant launched on the Nubia NaviX Ultra phone with the Screen Automation Execution Protocol, under which third-party apps opt in to permit AI-driven actions. The protocol makes the permission model an explicit product choice; the account is press reporting, not independently verified, and does not establish adoption or cross-app reliability.
Source 10: Doubao sets independent 12% transaction fee on in-assistant hotel orders
Xinlang Xinwen
Published: 2026-09-18
Reports that Doubao set an independent 12% transaction fee on in-assistant hotel orders, 11.4% service plus 0.6% payment, compared with 8% on Douyin's main site. This is the first independent pricing of AI-conversation traffic in China, valuing the assistant channel as distinct from the parent platform's traffic; it is a company-set policy, and adoption effects are unknown.
Source 11: DeepSeek API Change Log
DeepSeek
The company's change log records V4-Pro general availability in its app, web, and API on 13 August 2026, V4-Flash-Vision-Exp on 21 August, and V4.1-Flash on 10 September, with price reductions. The benchmark scores and capability descriptions are the company's own, not independent results, and say nothing about the consumer app's task outcomes.
Source 12: Kimi AI Assistant
Moonshot AI
Kimi's official page announces K3 and positions the assistant for agentic coding and knowledge work. This is company product positioning, not independent performance, adoption, retention, or outcome evidence.
Source 13: ????????? (Membership pricing and plans)
Moonshot AI
Kimi's official help-center pricing page describes four monthly subscription tiers (?49, ?99, ?199, and ?699, with lower monthly prices when billed annually), one shared usage-credit pool across the product's modes including K3 and Kimi Claw, a separate five-hour weekly cap for Kimi Code, and daily credit charges for a hosted Kimi Claw sandbox and a published Agent website. These are Moonshot's stated prices and mechanics, not independent measures of value, retention, or task outcomes.
Source 14: QuestMobile 2026 AI Platform Development Report
QuestMobile
Published: 2026-09-01
QuestMobile's September report on AI platforms, using July 2026 data. It describes office agents from Tencent, Alibaba, and ByteDance being brought together at group level. This is QuestMobile's analysis of activity and product structure, not a measure of task success or reliability.
Source 15: 关于 QClaw 停止运营的公告 (Notice that QClaw is ceasing operation)
Tencent
Published: 2026-09-24
Tencent's notice that QClaw stops operating at 00:00 on 24 December 2026. New sign-ups and subscription purchases stopped on 24 September, unused subscriptions can be refunded, and data is deleted from 24 March 2027. It offers migration of configuration, memory, agent persona, and conversation history to personal WorkBuddy accounts only. The reason given is a business adjustment. The notice is a company statement and does not describe how many people used the product.
Source 16: TraeWork
ByteDance
TRAE's official product page describes TraeWork as an AI work assistant for presentations, data analysis, research, document writing, and code development, with separate Work, Code, and Design modes synchronized across desktop and web. The trae.cn domain returned a 403 response during two retrievals; trae.ai loaded directly and is the first-party record used here. This is a company product description, not independent adoption, reliability, or task-outcome evidence.
Source 17: Baidu rebrands overseas AI office product as Kooko, surpasses 10 million users
EqualOcean
Published: 2026-09-22
Reports that Baidu rebranded its overseas AI office platform as Kooko at its September 21 AI Day. The company claims 10 million overseas users in the first year across roughly 200 countries, paying customers in 197 countries, and monthly recurring revenue growing 50% month over month. These are company claims reported through technology press, not independently verified measures.
Source 18: AutoClaw
Zhipu AI
Zhipu's official AutoClaw page describes tools for office documents, social-media content, financial and investment analysis, website and mini-program creation without code, browser automation, and messaging integration. This is a company product description, not independent adoption, reliability, or task-outcome evidence, and it does not establish an activity figure comparable with another product.
Source 19: 人工智能拟人化互动服务管理暂行办法 (Interim Measures for the Management of Anthropomorphic AI Interaction Services)
Ministry of Justice of the People's Republic of China
Published: 2026-04-10
Full text of the measures. Article 2 applies them to sustained emotional-interaction services that simulate a person's personality and manner, and excludes customer service, question answering, work assistants, education, and research services that involve no sustained emotional interaction. Article 32 sets the start date of 15 July 2026. The text does not say how any particular product is classified.
Source 20: CAC and four other bodies publish the Interim Measures for Anthropomorphic AI Interaction Services
Cyberspace Administration of China
Published: 2026-04-10
The regulator's announcement that the CAC, NDRC, MIIT, MPS, and SAMR jointly issued the measures on 10 April 2026, in force from 15 July 2026. It summarizes the duties for providers and does not classify any named product.
Source 21: QuestMobile 2026 AI Application Market Half-Year Report
QuestMobile
Published: 2026-07-14
QuestMobile's own report and the original of the figures that IT Home and TechNode relay. It dates its category total and app figures to June 2026, draws on data from May 2025 to June 2026 in its own databases, and defines an AI-native app as one built around AI. It does not say how users of more than one app are counted, and its disclaimer states that accuracy and completeness are not guaranteed. The activity figures do not establish market share, retention, task success, or user value.
Source 22: QuestMobile Report: China AI-Native App Activity
IT Home
Published: 2026-07-14
IT Home relays QuestMobile's 2026 first-half AI-app report, including consumer-app names and activity estimates. Its text dates the category total to May while the graphic caption describes a June app ranking. QuestMobile's own report dates both to June 2026, so IT Home's May date does not match the original (see the QuestMobile source record). Do not add the app figures, treat them as market shares, or use them as a ranking.
Source 23: QuestMobile: China's AI-Native Apps Reach 499 Million Monthly Active Users
TechNode
Published: 2026-07-14
TechNode summarizes QuestMobile's 2026 first-half AI-app report and repeats app-activity estimates also reported by IT Home. It is a secondary account of the same underlying study, not independent replication. It says the category total was reached by May; QuestMobile's own report dates it to June 2026. The figures do not establish market share, retention, reliable task completion, or user outcomes.
Source 24: China AI Rankings by App MAU, Issue 23 (June 2026)
AICPB
Published: 2026-07-06
A second provider's June 2026 monthly-active-user ranking, last updated on 6 July 2026. Its sizes for the same apps differ from QuestMobile's, and it includes products that QuestMobile's top ten leaves out, such as Quark and Baidu Netdisk. Its method is described only as a standardized global methodology. Neither provider's figures are an audited count, so the two cannot be reconciled or ranked against each other.
Source 25: QuestMobile 2026 Q1 AI Application Insight Report
QuestMobile
Published: 2026-04-21
QuestMobile reports that AI-native-app category monthly active users reached 440 million by March 2026, with Doubao at 345 million, Qianwen at 166 million, and DeepSeek at 127 million. It also gives quarterly average ?activity rates? of 33.5%, 17.1%, and 21% for those apps without defining the measure or relating it to monthly active users. Only the text summary was read; its charts did not render. The figures do not establish market share, retention, or task outcomes.