CtrlK
Doubao (豆包, literally "bean bun") is an artificial intelligence chatbot and virtual assistant developed by ByteDance, the parent company of TikTok. Launched in August 2023, Doubao has grown into China's most popular AI application by daily active users, surpassing 100 million DAU during the 2026 Lunar New Year period. The app is powered by ByteDance's proprietary Seed model family and is available on iOS, Android, web, and desktop platforms. Its international counterpart, originally released as Cici and later rebranded to Dola, serves users in select overseas markets.
Doubao offers a wide range of capabilities including text-based conversation, image generation (via the Seedream model), video generation (via the Seedance model), document analysis, voice interaction, and code generation. Through ByteDance's Volcengine cloud platform, the underlying Doubao models are also available as enterprise APIs, where aggressive pricing (as low as 99.3% below industry averages at launch) triggered a price war across China's AI sector in 2024.
ByteDance began internal testing of an AI chatbot in June 2023, initially under the codename "Yunque." The project was developed by ByteDance's Seed research team, which was established in early 2023 as the company's response to the rapid advancements in large language models following the release of ChatGPT in November 2022. ByteDance's earlier AI Lab, founded in 2016, was later fully merged into the Seed team to consolidate AI research efforts.
On August 17, 2023, the chatbot launched publicly under the name "Doubao" and began invitation-only testing in China. The underlying language model was initially called Skylark (also referred to as the "Lark Model") before being rebranded under the Doubao name. At the same time, ByteDance released an international version of the chatbot called Cici, targeting overseas markets through a Singapore-based subsidiary, SPRING (SG) PTE. LTD.
Doubao was among the first AI models in China to receive algorithm registration from Chinese regulators, allowing it to operate as a consumer-facing product under the country's generative AI regulations.
In May 2024, ByteDance formally launched the Doubao model family through its Volcengine cloud platform. The company set prices at 0.0008 yuan (roughly 0.00011 US cents) per 1,000 tokens for its enterprise-tier model. According to Tan Dai, president of Volcengine, this was 99.3% less than the industry average of approximately 12 Chinese cents per 1,000 tokens for models of equivalent specification. The aggressive pricing strategy immediately triggered a broader price war across China's enterprise large language model sector, forcing competitors to slash their own API fees.
By September 2024, Doubao became the first AI large model application in China to exceed 100 million total downloads. By November 2024, it had approximately 60 million monthly active users, making it China's most popular AI chatbot by that metric.
At the December 2024 Volcengine FORCE conference (Winter edition), ByteDance announced a comprehensive upgrade to the Doubao model family. Key releases included the Doubao General Model Pro (with 32% improved task-handling ability compared to May), a visual understanding model priced at just 0.003 yuan per 1,000 tokens (85% below the industry average), and a 3D generation model capable of producing high-fidelity 3D assets within one minute. ByteDance also announced that the Doubao video generation model would officially open for external services in January 2025.
The year 2025 saw Doubao's usage metrics accelerate dramatically. In January 2025, ByteDance released the Doubao-1.5-Pro model, a sparse Mixture of Experts (MoE) architecture with 20 billion activated parameters that matched the performance of a 140-billion-parameter dense model. The model came in two context-length configurations (32K and 256K tokens) and featured a "deep thinking" reasoning mode. Benchmarks showed it outperforming GPT-4o and Claude 3.5 Sonnet on knowledge, coding, reasoning, and Chinese language processing tasks.
In May 2025, ByteDance published the Seed1.5-VL technical report and released the model on Volcengine. This vision-language model, composed of a 532-million-parameter vision encoder and an MoE LLM with 20 billion active parameters, achieved state-of-the-art performance on 38 out of 60 public benchmarks. In agent-centric tasks such as GUI control and gameplay, Seed1.5-VL outperformed leading multimodal systems including OpenAI CUA and Claude 3.7.
By August 2025, according to QuestMobile's AI Application Industry Monthly Report, Doubao reached 157 million monthly active users with a 6.6% month-over-month growth rate, officially surpassing DeepSeek to claim the top position among native AI applications in China.
In September 2025, ByteDance released Doubao Vision, its first visual deep-thinking model, combining advanced visual understanding with reasoning capabilities and tool-calling functionality.
In October 2025, Volcengine launched Doubao 1.6-Vision, a multimodal model with enhanced tool-calling capabilities at 50% reduced cost compared to its predecessor.
In December 2025, at the Volcengine FORCE conference, ByteDance launched the Doubao Large Model 1.8 alongside the Seedance 1.5 Pro video generation model. The 1.8 model was optimized specifically for multimodal agent scenarios, with enhanced tool invocation, complex instruction compliance, and OS agent capabilities. It achieved world-leading performance on the BrowserComp general agent evaluation set. At this point, Doubao's daily average token consumption exceeded 50 trillion, representing a 417-fold increase since May 2024. More than 100 enterprise customers had each accumulated over one trillion tokens in usage.
Also in December 2025, ByteDance released Doubao-Seed-Code, which achieved state-of-the-art results on the SWE-Bench-Verified leaderboard while offering 62.7% lower costs compared to industry averages.
By late December 2025, Doubao surpassed 100 million daily active users, with 155 million weekly active users and approximately 172 million monthly active users. ByteDance noted that Doubao's user acquisition and marketing spend was the lowest among all ByteDance products that had ever reached the 100 million DAU milestone, attributing much of the growth to organic distribution through its ecosystem of apps including Douyin (the Chinese version of TikTok) and Toutiao.
On February 14, 2026, ByteDance released Doubao 2.0, powered by the new Seed 2.0 foundation model family. The release was positioned as an "agent-era" upgrade, shifting the product from question-and-answer interactions toward executing complex, multi-step real-world tasks autonomously.
The Seed 2.0 family includes four model variants: Seed 2.0 Pro for frontier reasoning and complex agents, Seed 2.0 Lite for general production workloads, Seed 2.0 Mini for high-throughput batch processing, and Seed 2.0 Code for software development (optimized for the TRAE IDE). The flagship Pro variant uses "System 2" thinking processes with advanced reasoning chains. ByteDance internally benchmarks Seed 2.0 Pro at parity with GPT-5.2 and Gemini 3 Pro on math, coding, and logical reasoning. Reported benchmark scores include 98.3 on AIME 2025, a 3020 Codeforces rating, and 89.5 on VideoMME for hour-long video processing.
Pricing for Seed 2.0 Pro was set at approximately $0.47 per million input tokens and $2.37 per million output tokens, roughly 3.7 times cheaper on input and 5.9 times cheaper on output than GPT-5.2.
Two days later, on February 16, 2026 (Lunar New Year's Eve), Doubao served as the exclusive AI cloud partner of the CCTV Spring Festival Gala, one of the most-watched television broadcasts in the world. During the live event, Doubao handled over 1.9 billion AI-related queries. The partnership included interactive features such as AI-generated New Year greetings, three rounds of lottery during the broadcast, and over 100,000 technology product giveaways (including DJI drones, Yuqi robots, and 3D printers). Doubao's DAU surged to over 100 million on that day, roughly four times its early-February levels.
As of March 2026, the competitive picture has continued to shift. Alibaba's revamped Quark AI assistant overtook Doubao in monthly active users, reaching approximately 150 million MAU compared to Doubao's roughly 100 million MAU, according to data from Aicpb.com. However, Doubao remains one of the most actively used AI products in China by engagement metrics.
Doubao is powered by ByteDance's Seed model family, developed by the company's Seed research team. The model lineage has evolved through several generations:
| Model | Release Date | Key Characteristics |
|---|---|---|
| Skylark (Yunque) | Mid-2023 | Original LLM, later renamed; powered the first version of Doubao |
| Doubao General Model Pro | May 2024 | First enterprise-grade model release on Volcengine |
| Doubao-1.5-Pro | January 2025 | Sparse MoE architecture, 20B active parameters, deep thinking mode |
| Seed1.5-VL | May 2025 | Vision-language model, 532M vision encoder + 20B MoE LLM, SOTA on 38/60 benchmarks |
| Doubao Vision | September 2025 | First visual deep-thinking model with tool-calling |
| Doubao 1.6-Vision | October 2025 | Multimodal model with enhanced tool-calling, 50% cost reduction |
| Doubao-Seed-Code | December 2025 | Coding-specialized model, SOTA on SWE-Bench-Verified |
| Seed 1.8 | December 2025 | Generalized agentic model with GUI, search, and code agent capabilities |
| Seed 2.0 (Pro/Lite/Mini/Code) | February 2026 | Agent-era foundation models, long-chain reasoning, GPT-5.2 parity claimed |
The Seed 1.8 model introduced native foundational vision capabilities that allow it to directly interact with graphical interfaces across desktop, web, and mobile environments. It supports tool invocation within its thinking mode (a technique also adopted by Claude Sonnet 4.5 and DeepSeek-V3.2) and can process up to 1,280 frames of video in a single pass.
The Doubao model family uses a sparse Mixture of Experts (MoE) architecture, where only a subset of parameters are activated for each input. For example, Doubao-1.5-Pro activates 20 billion parameters per inference pass while maintaining performance comparable to a 140-billion-parameter dense model. This approach allows ByteDance to offer high-performance AI at significantly reduced computational cost, which directly enables its aggressive pricing strategy.
The Seed 2.0 generation employs what ByteDance describes as "System 2" thinking, referring to multi-step deliberative reasoning processes (as opposed to fast, pattern-matching "System 1" responses). This architecture is designed for agentic workflows where the model must decompose complex objectives into executable sub-tasks and orchestrate multiple tools and API calls autonomously.
Doubao provides a broad set of AI-powered features across its consumer app and enterprise API offerings.
| Feature | Description | Underlying Model |
|---|---|---|
| Text Chat | Conversational AI assistant for Q&A, brainstorming, writing, and general knowledge | Seed 2.0 Pro / Lite / Mini |
| Deep Thinking | Extended reasoning mode for complex math, logic, and analysis problems | Seed 2.0 Pro (System 2 reasoning) |
| Image Generation | Text-to-image creation with support for Chinese aesthetics and 4K resolution | Seedream 4.5 (up to 5.0 Lite) |
| Video Generation | Text/image/video-to-video creation at native 2K resolution with audio sync | Seedance 2.0 |
| Visual Understanding | Image and video analysis, chart interpretation, OCR, and visual Q&A | Seed1.5-VL / Doubao Vision |
| Document Analysis | Upload and analyze PDFs, reports, and long-form documents; extract key points | Seed 2.0 with long-context support |
| Voice Interaction | Speech-to-text and text-to-speech with emotional expression and dialect support | Volcano Engine speech models |
| Code Generation | Programming assistance, debugging, and code review | Seed 2.0 Code / Doubao-Seed-Code |
| AI Agent Tasks | Multi-step autonomous task execution (travel planning, research, GUI automation) | Seed 1.8 / Seed 2.0 Pro |
| Web Search | Real-time information retrieval integrated into chat responses | Integrated search capabilities |
| Translation | Support for 28+ languages with quality comparable to GPT-4o | Doubao translation model |
Doubao's image generation is powered by the Seedream model family. Seedream 4.5, released in December 2025, generates native 4K resolution images from text descriptions and can process up to 14 reference images simultaneously. It achieves 94% accuracy on complex typography rendering and up to 90% character consistency across multiple generated images. The model has been specifically tuned for "Chinese aesthetics," producing artwork that reflects traditional Chinese artistic styles. The newer Seedream 5.0 Lite adds deep thinking and online search capabilities to the image generation pipeline.
The Seedance model family handles video generation within Doubao. Seedance 1.0 ranked first on both video generation leaderboards maintained by Artificial Analysis as of June 2025. The latest version, Seedance 2.0, released alongside Doubao 2.0 in February 2026, is the first model to accept four input modalities simultaneously: text, up to 9 images, up to 3 video clips, and up to 3 audio tracks. It produces cinematic-quality video at native 2K resolution (2048x1080 or 1080x2048) with synchronized audio, and offers 30% faster generation speeds compared to Seedance 1.5 Pro.
Doubao supports voice-based interaction through both its mobile and desktop apps. Users can tap a microphone icon to speak instead of typing, and the app responds with synthesized speech. By August 2024, Doubao's text-to-speech system had achieved precise emotional expression, and Volcengine announced plans for an end-to-end real-time voice model with capabilities including multi-character performance and dialect conversion.
Users can upload files (PDFs, images, spreadsheets) for analysis through the app's paperclip interface. Doubao can summarize long articles, extract key points from reports and meeting notes, interpret charts and graphs, and analyze video content. The Seed1.5-VL model provides the visual understanding backbone, excelling at complex reasoning, optical character recognition, image interpretation, and open-vocabulary detection.
Doubao is available across multiple platforms:
All platforms support account-based cloud synchronization, allowing users to access their conversations and settings across devices.
The international version of Doubao was originally launched as Cici in August 2023, targeting users outside mainland China. Cici closely resembled Doubao in interface design, supported text and voice interaction, included image generation and analysis, and allowed users to interact with intelligent agents created by other users.
In late 2025, ByteDance rebranded Cici to Dola. Existing users could log into Dola with the same account, with all previous chats and settings preserved. The app is developed and distributed globally by SPRING (SG) PTE. LTD., a Singapore-based subsidiary of ByteDance.
Dola is available in select markets including the UK, Mexico, and Spain, but is currently unavailable in the United States, Canada, and Australia. Some features available in the Chinese Doubao app, such as video generation, are blocked in the international version. However, ByteDance licenses both the Seedream and Seedance models to international platforms through the Volcengine API, allowing developers outside China to access these capabilities without downloading the consumer app.
Alongside Doubao, ByteDance operates Coze, a low-code platform for building custom AI chatbots and agents. Coze was introduced to the international market in January 2024 and is available in both a domestic version (known as Cozi in China) and an international version.
The platform provides a visual builder with drag-and-drop workflow design, plugin integrations (including Google Search, DALL-E 3, and CapCut), short-term and long-term memory systems, reusable UI cards, and deployment options to ecosystems such as Doubao, Feishu (ByteDance's enterprise collaboration tool), and Telegram. The domestic version runs on ByteDance's in-house Doubao models (with Doubao 1.5 Pro as the primary model), while the international version integrates external models such as GPT-4o and DALL-E.
In April 2025, ByteDance launched an internal beta for Coze Space, an AI agent collaboration platform that extends Coze's capabilities for team-based agent development.
In 2025, ByteDance open-sourced the core platform as Coze Studio, enabling developers to self-host and customize the bot-building infrastructure.
Doubao's models are commercially distributed through Volcengine (Volcano Engine), ByteDance's cloud computing and enterprise services platform. Volcengine offers the full Doubao model family through a Model-as-a-Service (MaaS) API, allowing enterprise customers to integrate Doubao's text, vision, code, and agent capabilities into their own products.
As of December 2025, daily token consumption on the Volcengine platform exceeded 50 trillion tokens, up from approximately 4 trillion one year earlier and just 120 billion at launch in May 2024 (a 417-fold increase). Over 100 enterprise customers have each accumulated more than one trillion tokens in total usage, spanning industries including e-commerce, finance, entertainment, and manufacturing.
In December 2025, Volcengine launched the Doubao Assistant API, making the core agent capabilities of the Doubao app (dialogue, reasoning, and search) directly available for enterprise integration. Future releases are planned to include multimodal understanding, in-depth research, content creation, and video call capabilities through the API.
ByteDance has consistently positioned Doubao as one of the most cost-effective AI model offerings globally:
| Model / Period | Pricing | Comparison |
|---|---|---|
| Doubao enterprise model (May 2024 launch) | 0.0008 yuan per 1,000 tokens | 99.3% below industry average of 12 yuan cents per 1,000 tokens |
| Doubao-1.5-Pro (January 2025) | Comparable to GPT-4o quality | Approximately 50x cheaper than GPT-4o |
| Doubao Visual Understanding (December 2024) | 0.003 yuan per 1,000 tokens | 85% below industry average |
| Seed 2.0 Pro (February 2026) | $0.47/M input, $2.37/M output tokens | 3.7x cheaper input, 5.9x cheaper output vs. GPT-5.2 |
This pricing strategy relies on the computational efficiency of the MoE architecture (activating only a fraction of total parameters per query) and ByteDance's massive infrastructure investments.
ByteDance has committed substantial capital to AI infrastructure. Reports indicate the company planned approximately $20 billion in capital expenditure for 2025, with a significant portion directed toward AI infrastructure development. For 2026, ByteDance has preliminarily budgeted RMB 160 billion (approximately $23 billion) in capital expenditures, with roughly half allocated to procuring advanced semiconductors for AI model development. The company has faced challenges related to U.S. export controls on advanced GPUs, reportedly setting aside approximately $7 billion to rent NVIDIA GPUs through cloud services outside China as a workaround.
Doubao operates in an intensely competitive Chinese AI chatbot market. Its primary competitors include:
| Product | Developer | Notable Metrics (as of early 2026) |
|---|---|---|
| Doubao | ByteDance | 100M+ DAU (Feb 2026), 172M MAU (late 2025) |
| DeepSeek | DeepSeek (High-Flyer) | ~77-145M MAU (varies by period) |
| Ernie Bot (Wenxin Yiyan) | Baidu | 300M+ total users; became free in April 2025 |
| Tongyi Qianwen (Qwen) | Alibaba | ~150M MAU via Qwen models |
| Quark AI | Alibaba | ~150M MAU (March 2026, #1 by MAU) |
| Kimi | Moonshot AI | Enhanced coding and reasoning with Kimi K2 (mid-2025) |
| Zhipu AI (ChatGLM) | Zhipu AI | Enterprise-focused, open-source GLM models |
Doubao's competitive advantages stem from ByteDance's distribution network (leveraging Douyin and Toutiao for organic user acquisition), aggressive pricing that undercuts competitors by an order of magnitude, and the breadth of its multimodal capabilities (text, image, video, voice, and code in a single app). The integration of Seedream for image generation and Seedance for video creation gives Doubao a content-creation edge that aligns with ByteDance's core competency in short-form video.
The competitive rankings have been fluid. While Doubao led in DAU and MAU for most of 2025, Alibaba's Quark overtook it in MAU by March 2026 after Alibaba repositioned Quark from a cloud storage and search app into an "AI super assistant" powered by Qwen reasoning models.
| Date | Milestone |
|---|---|
| June 2023 | ByteDance begins internal testing of AI chatbot (codenamed Yunque) |
| August 2023 | Doubao launches publicly; international version Cici released |
| May 2024 | Doubao model family launched on Volcengine at 99.3% below industry pricing |
| September 2024 | First Chinese AI app to exceed 100M total downloads |
| November 2024 | Becomes China's most popular AI chatbot (~60M MAU) |
| December 2024 | Visual understanding and 3D generation models released at FORCE conference |
| January 2025 | Doubao-1.5-Pro released with deep thinking mode; video generation opens to public |
| May 2025 | Seed1.5-VL vision-language model released (SOTA on 38/60 benchmarks) |
| August 2025 | 157M MAU, surpasses DeepSeek as top native AI app in China |
| December 2025 | Seed 1.8 agentic model and Seedance 1.5 Pro released; 50T daily tokens; 100M DAU reached |
| February 2026 | Seed 2.0 / Doubao 2.0 released; CCTV Spring Festival Gala partnership; 1.9B queries on Lunar New Year's Eve |
| March 2026 | Alibaba's Quark overtakes Doubao in MAU rankings |