Doubao Mobile 3.0

• 1. overview (865 chars)

  Doubao Mobile 3.0 is a major version upgrade of ByteDance's AI assistant "Doubao" completed in 2026. The core is to push the past "chat robot" into the two directions of "personal digital assistant" and "productivity tool". This version replaces the underlying large model with Seed 2.0, opens up the expert mode for free, launches real-time online reasoning of "thinking and searching", and adds a new in-app shopping closed loop of "Beanbao helps you choose". At the same time, it has made a lightweight interface revision for middle-aged and elderly users and zero-experience users. Relying on Byte Ecosystem’s traffic infusion and all-scenario free strategy, Doubao has led the domestic AI native app market with approximately 380 million monthly active users, exceeding the second and third place combined in scale. However, high monthly activity also brings heavy computing power costs. Doubao is testing the boundaries of C-side charging. Whether commercialization can be successful is still the biggest suspense.

• 2. Product background (1068 chars)

  Doubao was developed by ByteDance and is currently the largest AI native application in China. According to QuestMobile data, as of March 2026, Doubao’s monthly active users were approximately 345 million, which had risen to approximately 382 million by June, a year-on-year increase of 172.1%. It is the only product that remains firmly in the first tier. The second place is Qianwen with approximately 167 million monthly active users, DeepSeek with approximately 130 million, Yuanbao with approximately 49.84 million, and Ant Afu with approximately 28.97 million. Doubaojia’s monthly active users exceeded the second and third place combined, and the gap further widened to more than 200 million in the second quarter.   Beanbao’s lead is no accident. It relies on the built-in AI interface of super traffic portals such as Douyin and Toutiao to import a large number of ordinary users who watch videos, and then retains them with "all daily functions are free". The user stickiness data is also outstanding: the average user opens 76.7 times per month and uses 143.7 minutes, both indicators are higher than the industry average. The essence of this 3.0 upgrade is that Byte wants to transform these massive users from "casual questions" to "inseparable daily tools."

• 3. Product functions and usage experience (2683 chars)

  The capability upgrades in version 3.0 focus on four main lines.   The underlying model was replaced with Seed 2.0, forming a model family of Mini (fast), Lite (daily), Pro (professional), and Code (programming). The inference efficiency increased by about 43%, supported millions of token contexts, and made long documents and multi-round conversations more stable. What attracts the most attention is that the "Expert Mode" is free and open, giving everyone the Pro-level capabilities that were only available for payment in the past. It is suitable for writing papers, complex analysis, and intensive reading of long documents. Actual measurement and comparison of a 127-page policy document showed that there is almost no difference in key information recall rate and logical chain integrity between the free expert mode and the original paid version. The only difference is that the response speed is about 0.6 seconds slower. Behind this is the computing power balance of "dynamic resource pool hierarchical scheduling + intelligent downgrade".   "Search while thinking" changes the traditional "search first and think later" to reasoning while supplementing information. When encountering complex problems, it will automatically break down sub-tasks and verify them step by step. It has a built-in authoritative source whitelist, time-limited circuit breaker and cross-validation mechanism. When checking policies, market conditions, weather, and news, it will pull government official websites, central media and other sources, and mark the specific source and update time at the end of the answer. The value of this design is not in speed, but in "credibility" - encapsulating the authenticity of information into every question.   In terms of multi-modality, it is equipped with Seedance 2.0 video generation (2K video in one sentence, multi-lens audio and video synchronization), SeedEdit 3.0 intelligent P-picture (partial editing, style cloning), as well as capabilities such as photo recognition and long video understanding. The office scene has enhanced voice transcription (dialect, weak network optimization), three-minute extraction of hundreds of pages of PDF, one-click export of chat records, and cross-application automation (one-stop from writing to finishing PPT to email). In June 2026, "Task Mode" was also launched separately, focusing on multi-round search, deep reasoning, browser automation and multi-modal content generation, marking the evolution of Doubao from a text tool to an AI Agent that can handle complex workflows.   Another obvious move in 3.0 is the lightweighting of the interface. The homepage only retains four major sections: life tools, graphic and text creation, smart Q&A, and local services. Promotional pop-ups are closed by default. Users over 60 years old automatically enter "Elder Simple Mode" after logging in, with enlarged fonts, bold icons, and fixed voice buttons at the bottom. Full-scenario offline speech recognition covers 28 dialects across the country. These changes are obviously aimed at middle-aged and elderly users and zero-based users, and they want to further penetrate the market.

• 4. Pricing and business model (1459 chars)

  The core strategy of Doubao is that "most livelihood, office, and learning tools are permanently free." There are no mandatory memberships, no watermarks, and unlimited times. This is also the key to widening the gap between users. But the huge free disk is eating back profits: unlike the traditional Internet, where the marginal cost is diminishing, every interaction with AI generates real computing power overhead for GPU inference and model operations, and the cost rises linearly with usage. CITIC Construction Investment's research report shows that as of March 2026, the average daily Token usage of the Doubao large model has exceeded 120 trillion, doubling in three months and 1,000 times higher than when it was released in May 2024; among them, video generation is the scene with the highest single task consumption, and a 10-second 1080p video consumes approximately 350,000 Tokens.   In this context, Doubao began to test the waters of C-end charging, which is regarded as a landmark event in the commercialization of domestic AI assistants. In comparison, Kimi has four levels of charging (continuous monthly subscription is 49/99/199/699 yuan), Zhipu Qingyan has VIP and SVIP (monthly card is 79/229 yuan), and bean bags are not the most expensive, but as the industry’s first charging exploration, it has more benchmark significance. At the same time, ByteDance is also using "Doubao to help you choose" to build an AI e-commerce closed loop - in a word, the product card needs to be directly issued, and browsing, selecting specifications, and payment (binding with Douyin payment) can be completed within the app without having to jump to Douyin, trying to move from traffic monetization to transaction monetization.

• 5. User reviews (1048 chars)

  Positive feedback focuses on “free and playable” and “low threshold to get started”. After testing it for three weeks, a long-term user said that from writing a proposal report, checking the medical insurance details to making a product plan and cutting a short video, there was no need to open other apps during the entire process, and they commented that it is like a "life-level smart terminal like a rice cooker." Middle-aged and elderly users have a high degree of recognition of the "Elder Simple Mode + Dialect Voice + Screen Sharing" and believe that it is finally an AI that "they can use independently." People's livelihood functions such as accounting, recipes, local convenient inquiries, and homework photos and lectures are frequently used by ordinary users.   There are two main points of negative and reserved opinions. The first is privacy concerns. Officials also remind not to upload sensitive information such as complete documents, bank card numbers, etc. Second, although the free expert model is strong, there are perceptible differences in slow response and degraded computing power during peak periods; some advanced users are worried that as charging advances, the existing free capabilities will be gradually narrowed.

• 6. Industry evaluation (1217 chars)

  The industry generally believes that Doubao represents the trend of domestic large-scale model applications from "pan-entertainment dialogue" to "high-value productivity tools." Chen Yan, research director of QuestMobile, pointed out that after the Spring Festival red envelope-style pulse of new users has subsided, the real test of product acceptance will be whether the monthly activity of top applications can be stabilized. Doubao’s monthly activity still increased by 22.3% in the second quarter, showing strong retention ability. Analysts also reminded that traffic is accelerating towards the top. The top two products account for most of the users on the list, making it increasingly difficult for small and medium-sized AI tools to break through.   However, the media also directly discussed the cost pressure on bytes. According to reports, ByteDance’s net profit in 2025 will decline by more than 70% year-on-year, which is alleged to be related to the substantial increase in AI full-chain investment and high-level computing power purchases in the second half of 2025 (Byte responded that this caliber is affected by the cost of preferred shares and options and does not reflect actual operations). The H100 one-year GPU rental price increased from US$1.70 per hour in October 2025 to US$2.35 in March 2026, an increase of nearly 40%, further amplifying the structural problem of Doubao "the more users, the higher the cost."

• 7. Controversies and Risks (1057 chars)

  The biggest controversy is the sustainability of the business model. Doubao gained the largest scale by offering free services, but it also put itself on the balance between "huge users" and "heavy costs". Charging too quickly may lose price-sensitive sinking users, charging too slowly may make it difficult to cover the soaring computing power bills. This is described by the industry as "a divisive experiment in the commercialization of AI in the C-end market."   The second is content compliance and information security risks. Doubao is deeply involved in people's livelihood information such as government affairs, medical care, education, and finance. Once the timeliness or source is wrong, it may mislead users. Although the official has used source whitelists and cross-verification to ensure the bottom line, the boundaries still need to be continuously invested in maintaining. Commercial copywriting also has compliance risks such as extreme advertising words, which the product uses one-click compliance detection to avoid. In addition, the pressure on privacy and data security brought by the large number of pictures, documents, and voices uploaded by users is a long-term problem that this type of national-level AI cannot avoid.

• 8. Applicable groups and usage suggestions (1001 chars)

  Doubao Mobile 3.0 is suitable for ordinary users who pursue "one app to solve 80% of daily trivial matters" - students, office workers, moms, small businesses, and middle-aged and elderly people. It is especially suitable. It is free, has dialect voice, and has an elder mode so that people with no basic knowledge can get started quickly. For those who need free long document analysis, real-time trusted information retrieval, lightweight video editing and image and text creation, it is very cost-effective.   It is not suitable for professional and heavy users who are extremely sensitive to data privacy, or who require stable top-end computing power and a non-degrading experience. Such needs may require a paid version or more vertical tools. Users who want to do serious programming and professional design should also use Doubao as an auxiliary rather than the main force. As an alternative, you can look at Qianwen, DeepSeek, and Yuanbao for general dialogue. For programming, you can look at specialized AI programming tools. The gap between each company in long document and retrieval scenarios is narrowing. You can choose according to the free quota and the quality of the source.

• 9. Summarize (510 chars)

  Doubao Mobile 3.0 uses "free in all scenarios + ecological perfusion + sinking of capabilities" to build itself into a national-level AI entrance, which is unrivaled in terms of scale; but its real test is not in function, but in how to turn 380 million monthly users from a cost burden into sustainable business value. This version is a "both offensive and defensive" answer sheet submitted by Byte - attacking on productivity and e-commerce closed loop, defending on sinking and retention. Whether it can achieve the balance between charging and computing power will determine the future ceiling of Doubao.

No in-depth report yet

• 

  MarkLee_X

  —

  专家模式这次居然免费开放，我特意拿它和之前订阅的付费版做对比，同一份127页的政策申报指南，让它做摘要和条款风险识别，两边关键信息召回率几乎一模一样，逻辑链也都能追溯到具体条款编号，唯一能感觉到的区别就是免费版慢了一秒多，晚高峰测了十次全部成功，这波是真给普通人省钱了。

---

• 

  JCooper007275

  —

  3.8亿月活，断层第一不是白叫的。

---

• 

  菊花_10

  —

  长辈简易模式给我妈用真的绝，字体放大两倍，语音按钮固定在页面最底下不用找，四川话也能精准识别，她58岁自己三天就学会了在线挂号、查检验报告，还给我侄子录英语发音，以前这些都得我手把手教半天，现在完全不用管了。

---

• 

  TheCeylanBeşerler_2024

  —

  边想边搜这个我最爱，查政策会直接标出处和更新时间，不用自己再去甄别真假，比以前满网页翻靠谱多了。

---

• 

  Judy_ScottQ

  —

  免费是真香，但我有点担心，等以后收费了这些功能会不会又收回去。

---

• 

  BettyTorresSr

  —

  拍照记账那个功能太实用了，超市小票一拍自动分类，月底还生成省钱建议，我这种懒人终于能坚持记账了。

---

• 

  天涯616

  —

  卡是不卡，就是晚上高峰偶尔要多等几秒，能理解，毕竟这么多人用。

---

• 

  zhy46z03_h5

  —

  任务模式试了一下真有点东西，让它做一份行业市场分析，它自己把大目标拆成好几个子任务，联网搜资料、调浏览器抓实时信息，最后直接给我出了网页、文档报告和PPT大纲，全程基本没怎么插手，从聊天工具变成能干活的AI Agent，这个方向是真选对了。

---

• 

  夏风279

  —

  回不去了。

---

• 

  Alyssa71_io

  —

  作业拍题精讲对家长友好，它不直接给答案，先讲知识点引导孩子想，这点比某些搜题软件强多了，不怕孩子抄。

---

• 

  Andrea.RichardsonJr

  —

  说句实话，隐私这块我还是有点顾虑，官方也提示别传身份证银行卡，日常用还是收着点好。

---

• 

  happyfrog634

  —

  豆包帮你选真的把买东西这件事简化了，说个预算和需求直接出对比表，还带实时价和优惠券，省了我刷好几个平台的时间。

---

• 

  Donald877

  —

  免费的，功能还这么全，图片视频文档全都能搞，字节是真舍得砸钱。

---

• 

  OMillerJr314

  —

  方言识别是真的强，我用四川话说买菜账它都能算对，离线也能用，弱网环境下也没掉链子。

---

• 

  Roger_CruzZ3

  —

  看到新闻说字节去年利润跌了七成，据说一大半是砸在AI算力上，豆包这么大用户量，每次对话都是真金白银的GPU推理成本，不像传统App用户越多越赚，它是用户越多成本越高，收费迟早的事，就看怎么收不把下沉用户吓跑。

---

• 

  4vd_491

  —

  界面清爽多了，弹窗广告默认关掉，首页就四个大分区，长辈也不会点晕，这次改版走心了。

---

• 

  Linda_Nguyen209

  —

  Seed 2.0之后长文档和多轮对话稳了很多，以前聊到后面它会忘前面，现在百万上下文基本不掉记忆。

---

• 

  PoolRunnerjchuster

  —

  太强了。

No reviews yet