CtrlK
ByteDance AI refers to the artificial intelligence research, products, and infrastructure developed by ByteDance, the Chinese technology company best known as the parent of TikTok. Founded in 2012 by Zhang Yiming and Liang Rubo, ByteDance built its business on AI-powered content recommendation from the very beginning, starting with the Toutiao news aggregation app. The company has since expanded into large language models, generative media, cloud AI services, and autonomous agents, making it one of the most significant players in the global AI landscape.
As of early 2026, ByteDance operates the most popular AI chatbot in China (Doubao), runs the leading large model cloud service platform by market share (Volcano Engine), and invests tens of billions of dollars annually in AI infrastructure. Its Seed research team develops foundation models that compete with those from OpenAI, Google, and Anthropic on international benchmarks.
ByteDance was founded in March 2012 in Beijing by Zhang Yiming and Liang Rubo, who had been college roommates at Nankai University. Zhang's original vision was to use machine learning algorithms to push relevant content to smartphone users rather than requiring them to search for it. At the time, most venture capitalists rejected the idea, and the company secured its first investment from Susquehanna International Group.
The company's first product, Toutiao, launched in August 2012 as a news aggregation app. Toutiao used AI to track user behavior, including reading duration, scroll patterns, and content preferences, to build detailed profiles and serve increasingly personalized recommendations. Within two years, the app attracted more than 13 million daily active users. Sequoia Capital led a $100 million investment round in 2014.
ByteDance launched Douyin (the Chinese version of TikTok) in September 2016 and expanded internationally as TikTok in 2017. The company applied the same recommendation engine principles from Toutiao to short-form video, which proved enormously successful worldwide.
Liang Rubo succeeded Zhang Yiming as CEO in 2021, while Zhang transitioned to a role focused on long-term strategy. As of August 2025, ByteDance's valuation stood at over $330 billion based on its latest employee share buyback program. The company's 2025 revenue was estimated at approximately $186 billion, surpassing Meta in quarterly revenue for the first time in early 2025.
The Seed team is ByteDance's core AI research division, responsible for developing the company's foundation models and advancing fundamental research. The team's name reflects its mission: to plant the "seeds" of next-generation intelligence.
In 2025, ByteDance completed a major restructuring of its AI organization. The company's separate AI Lab was fully merged into Seed, consolidating research directions including AI for Science, Robotics, and Responsible AI under a single umbrella. The restructuring brought in Wu Yonghui, a former Vice President at Google DeepMind, as the lead for foundational research. Wu reports directly to CEO Liang Rubo.
The Seed organization is divided into several key groups:
The Seed team works across a broad range of AI research domains:
ByteDance has published extensively at top AI conferences. The company had over 50 research papers accepted at NeurIPS in recent years, and 25 papers accepted at ICML 2025. A joint paper with Peking University on visual autoregressive modeling received an Outstanding Paper award at NeurIPS.
The Seed model family is ByteDance's lineup of proprietary and open-source foundation models. The family has grown rapidly since 2024, spanning general-purpose LLMs, vision-language models, code-specialized models, and mathematical reasoning systems.
Seed2.0 is ByteDance's flagship model series, released on February 14, 2026. It comes in three sizes: Pro, Lite, and Mini. ByteDance claims Seed2.0 Pro matches or exceeds GPT-4.5 and Gemini 2.5 Pro on many benchmarks at roughly one-tenth the cost.
Key benchmark results for Seed2.0 Pro include:
| Benchmark | Seed2.0 Pro Score | Notes |
|---|---|---|
| AIME25 (Math) | 98.3 | Competitive math problems |
| GPQA Diamond | 88.9 | Graduate-level science questions |
| Codeforces Rating | 3020 | Competitive programming |
| LiveCodeBench v6 | 87.8 | Real-world coding tasks |
| SWE-Bench Verified | 76.5 | Software engineering tasks |
| Terminal Bench 2.0 | 55.8% | Agentic capabilities |
| VideoMME | 89.5% | Video understanding |
| LMSYS Chatbot Arena | 6th overall | Community-voted text quality |
| LMSYS Vision Arena | 3rd overall | Community-voted vision quality |
Seed2.0 also earned gold medals on ICPC, IMO, and CMO competitive mathematics and programming challenges.
Seed1.5-VL is a vision-language foundation model released on May 12, 2025. It uses a 532-million-parameter vision encoder paired with a Mixture of Experts (MoE) LLM with 20 billion active parameters. The model achieved state-of-the-art performance on 38 out of 60 public vision-language benchmarks at the time of release.
Seed-Coder is a family of 8-billion-parameter open-source code LLMs, including base, instruction-tuned, and reasoning variants. The models were trained on 6 trillion tokens with a focus on reducing human involvement in code data curation, instead relying on model-driven data pipelines.
Seed-OSS is a series of open-source general-purpose LLMs designed for international use. The flagship Seed-OSS-36B has 36 billion parameters and supports a 512K-token context window. The models and code were published on GitHub and Hugging Face.
Seed-Prover is ByteDance's formal mathematical theorem-proving system. It uses a lemma-style whole-proof reasoning approach that iteratively refines proofs based on Lean compiler feedback. Seed-Prover 1.5, released in late 2025, solved 5 out of 6 IMO 2025 problems within 16.5 hours, achieving a gold-medal-equivalent score. On the MiniF2F benchmark, it reached a 99.6% proof rate.
| Model | Type | Parameters | Release | Key Feature |
|---|---|---|---|---|
| Seed2.0 Pro | General LLM | Not disclosed | Feb 2026 | Flagship, agent-capable |
| Seed2.0 Lite | General LLM | Not disclosed | Feb 2026 | Balanced cost/performance |
| Seed2.0 Mini | General LLM | Not disclosed | Feb 2026 | Lightweight, fast inference |
| Seed1.5-VL | Vision-Language | 20B active (MoE) | May 2025 | 38/60 SOTA benchmarks |
| Seed-OSS-36B | Open-source LLM | 36B | Aug 2025 | 512K context window |
| Seed-Coder | Code LLM | 8B | Jun 2025 | Model-centric code data |
| Seed-Prover 1.5 | Math Reasoning | Not disclosed | Late 2025 | IMO gold-medal level |
| Seed-X | Translation LLM | 7B | 2025 | Multilingual translation |
| Seedream 5.0 | Image Generation | Not disclosed | 2026 | Text-to-image |
| Seedance 2.0 | Video Generation | Not disclosed | Feb 2026 | Text/image-to-video |
Doubao is ByteDance's consumer-facing AI assistant and chatbot, launched in August 2023. The name "Doubao" (豆包) translates roughly to "bean bun," a common Chinese snack. It was initially available through invitation-only testing before opening to the general public.
Doubao is a distinct product from Coze, ByteDance's AI agent-building platform. While Coze allows developers to build custom AI agents using workflow automation, Doubao is the primary consumer chatbot. The international version of Doubao is called Cici, which launched in August 2023 and is available in markets including the United Kingdom, Mexico, Indonesia, Malaysia, and the Philippines. Cici is not available in the United States or mainland China.
Doubao's growth trajectory has been remarkable. By November 2024, it had become China's most popular AI chatbot with approximately 60 million monthly active users. During the 2026 Lunar New Year holiday, Doubao's daily active users surged to over 100 million on February 16, 2026, roughly four times its levels from early February. This surge was driven in part by ByteDance's partnership with CCTV's Spring Festival Gala, one of China's most-watched television programs. Doubao handled over 1.9 billion AI-related queries during the broadcast.
For context, during the same Lunar New Year period, Alibaba's Qwen chatbot peaked at 30 million DAU despite a 3-billion-yuan promotional campaign, and Tencent's Yuanbao chatbot reached 50 million DAU with a 1-billion-yuan coupon giveaway.
Doubao supports text chat, voice conversation, image generation and analysis, document processing, and code assistance. It is powered by ByteDance's Seed model family. The chatbot is integrated across ByteDance's product ecosystem, including Douyin (the Chinese version of TikTok).
ByteDance has pursued an aggressive pricing strategy for Doubao's API access. When the Doubao Pro model launched for enterprise use in May 2024, it was priced at 0.0008 yuan per 1,000 input tokens, roughly 99.3% less than GPT-4 at the time. This pricing sparked a price war among Chinese AI companies.
As of 2025, Doubao-1.5 Pro-32k costs 0.8 yuan (approximately $0.11) per million input tokens, with output tokens priced at 2 yuan per million. The lightweight Doubao-1.5 Lite-32k is even cheaper at 0.3 yuan per million input tokens. These prices are roughly 70% below industry averages in China and a fraction of what Western providers charge.
| Model Tier | Input Price (per 1M tokens) | Output Price (per 1M tokens) |
|---|---|---|
| Doubao-1.5 Pro-32k | 0.8 yuan (~$0.11) | 2.0 yuan (~$0.27) |
| Doubao-1.5 Lite-32k | 0.3 yuan (~$0.04) | 0.6 yuan (~$0.08) |
| Seed2.0 Pro (API) | ~$0.47 | ~$2.37 |
Volcengine, known internationally as Volcano Engine, is ByteDance's cloud computing and enterprise services platform. Launched as a public cloud service in December 2021, it opens ByteDance's internal technology capabilities to external businesses.
Volcano Ark is ByteDance's Model-as-a-Service (MaaS) platform within Volcengine. It provides enterprises with model fine-tuning, evaluation, inference, and deployment services. The platform hosts ByteDance's own Doubao models alongside third-party models.
According to IDC's report on China's public cloud large model service market, Volcengine commanded a 46.4% market share in 2024, measured by API call volume. This exceeded the combined market share of the second and third-place providers. By the first half of 2025, Volcengine's share had grown to 49.2%, with Alibaba Cloud following at 27%.
| Provider | Market Share (H1 2025) |
|---|---|
| Volcengine (ByteDance) | 49.2% |
| Alibaba Cloud | 27.0% |
| Tencent Cloud | ~7% |
| Huawei Cloud | ~6% |
| Baidu AI Cloud | ~6.1% |
Volcengine generated over 12 billion yuan in revenue in 2024 and targeted more than 25 billion yuan for 2025.
Beyond large model APIs, Volcengine offers a comprehensive suite of cloud AI services:
TikTok's recommendation algorithm is widely regarded as one of the most influential AI systems in consumer technology. MIT Technology Review named it one of its 10 Breakthrough Technologies of 2021.
The recommendation system uses a combination of collaborative filtering and content-based filtering. At its core is a system called Monolith, a real-time recommendation engine with a collisionless embedding table designed to handle dynamic, sparse data while adapting to user feedback in real time.
When a user uploads a video, TikTok's pipeline analyzes it using natural language processing and computer vision technology. The system extracts information from the video content, audio, captions, and metadata to classify and understand the material.
For each user, the algorithm tracks interactions including:
Unlike algorithms on many other social media platforms, TikTok's system prioritizes content relevance based on individual behavior rather than follower count or account popularity. This means new creators can achieve viral reach without an existing audience, which was a key differentiator in TikTok's rise.
The recommendation engine developed for Toutiao and later refined for Douyin and TikTok represents ByteDance's foundational AI competency. The company's expertise in building large-scale recommendation systems directly informed its later work in training large language models and generative AI, since both domains require processing massive datasets, training at scale, and optimizing for user engagement.
Beyond text-based models, ByteDance has invested heavily in generative media, including image, video, and audio generation.
Seedance 2.0 is ByteDance's AI video generation model, released in February 2026. It supports text-to-video and image-to-video generation with cinematic quality. The model is integrated into ByteDance's Dreamina creative platform (known as Jimeng in China). Independent evaluations have found it competitive with or superior to OpenAI's Sora and Kuaishou's Kling in areas such as physics accuracy, lip-sync quality, and prompt adherence.
OmniHuman-1, released in early 2025, generates realistic human videos from a single photograph and motion signals. The system can produce full-body videos of people speaking, singing, and moving naturally, trained on 18,700 hours of human motion data. It uses a Diffusion Transformer architecture and can accept audio, video, or combined motion inputs.
Seedream 5.0 is ByteDance's text-to-image generation model, available through the Dreamina platform. It offers features including prompt-based generation, reference image styling, brush-based editing, and viewpoint control.
| Product | Category | Description | Availability |
|---|---|---|---|
| Doubao | AI Chatbot | Consumer AI assistant powered by Seed models | China |
| Cici | AI Chatbot | International version of Doubao | UK, Mexico, Southeast Asia |
| Coze | Agent Platform | No-code AI agent builder with workflow automation | Global |
| Dreamina (Jimeng) | Creative Suite | AI image, video, and music generation platform | China, Global |
| Volcano Ark | MaaS Platform | Enterprise model fine-tuning and inference | Global |
| CapCut | Video Editing | AI-powered video editing with effects and templates | Global |
| Toutiao | News Aggregation | AI-driven personalized news feed | China |
| TikTok / Douyin | Short Video | AI-powered content recommendation platform | Global / China |
ByteDance's ownership of TikTok has been the subject of sustained US government scrutiny since 2019. The Committee on Foreign Investment in the United States (CFIUS) launched an investigation into ByteDance on November 1, 2019, examining the company's 2017 acquisition of Musical.ly (which became TikTok).
The primary concerns centered on China's national security laws, which give the Chinese government the power to compel Chinese companies to hand over user data. TikTok collects extensive user information, including IP addresses, geolocation data, device identifiers, browsing history, and cookies. US officials worried that this data could be accessed by the Chinese government for intelligence purposes or that the platform could be used for propaganda or influence operations.
In August 2020, the Trump administration ordered ByteDance to divest TikTok. TikTok and its users challenged the order in court, and federal judges issued preliminary injunctions blocking enforcement. The Biden administration withdrew the executive orders in 2021.
However, concerns persisted. On April 24, 2024, President Biden signed the Protecting Americans from Foreign Adversary Controlled Applications Act (PAFACA) into law. The legislation required ByteDance to divest its interest in TikTok's US operations by January 19, 2025, or face a nationwide ban.
TikTok challenged the law, but the DC Circuit Court of Appeals upheld it. On January 17, 2025, the US Supreme Court unanimously rejected TikTok's appeal in an unsigned opinion with no dissents, ruling that the law did not violate the First Amendment.
On December 18, 2025, TikTok signed binding agreements with a consortium led by Oracle Corporation, Silver Lake, and MGX (an Abu Dhabi-based investment firm) to form TikTok USDS Joint Venture LLC. The deal closed on January 22, 2026.
The ownership structure of the new joint venture is:
| Stakeholder | Ownership Share |
|---|---|
| New investors (Oracle, Silver Lake, MGX at 15% each) | 50.0% |
| Existing ByteDance investors (affiliates) | 30.1% |
| ByteDance | 19.9% |
Oracle serves as the designated security partner, managing data audits, algorithm replication, and compliance with US national security regulations. The joint venture covers not only TikTok but also ByteDance's Lemon8 and CapCut applications in the US market.
ByteDance has committed massive capital expenditure to AI infrastructure. In 2025, the company spent approximately 150 billion yuan ($21 billion) on AI infrastructure. For 2026, ByteDance has budgeted approximately 160 billion yuan ($23 billion), with about half allocated to AI chips and semiconductors.
Specific investment plans include:
These investments place ByteDance among the world's largest purchasers of AI computing hardware, alongside Microsoft, Meta, Google, and Amazon.
ByteDance competes with several major Chinese technology companies in the AI space. The competition spans model development, cloud services, consumer applications, and hardware.
| Company | Primary AI Chatbot | Cloud AI Platform | Notable Models |
|---|---|---|---|
| ByteDance | Doubao | Volcengine / Volcano Ark | Seed2.0, Seedance, Seedream |
| Alibaba | Tongyi Qianwen | Alibaba Cloud | Qwen series |
| Baidu | Ernie Bot | Baidu AI Cloud | ERNIE series |
| Tencent | Yuanbao | Tencent Cloud | Hunyuan series |
| DeepSeek | DeepSeek Chat | N/A | DeepSeek V3, R1 |
ByteDance holds a clear lead in consumer AI adoption with Doubao's 100 million DAU and in cloud AI market share through Volcengine. However, Alibaba leads in overall cloud infrastructure market share at 35.8%, compared to Volcengine's 14.8% for general cloud services (as distinct from model-specific API calls).
The competition intensified during the 2026 Lunar New Year, when all major Chinese tech companies launched aggressive promotional campaigns for their AI assistants. ByteDance distributed 100,000 prizes, including luxury cars, during the CCTV Spring Festival Gala to promote Doubao. Alibaba spent 3 billion yuan promoting Qwen, while Tencent ran a 1-billion-yuan coupon campaign for Yuanbao.
In the talent war, ByteDance has been the most aggressive recruiter. Out of approximately 10,000 job openings in early 2026, 2,353 were focused on AI, compared to 357 AI roles at Baidu and 332 at Tencent.
| Date | Event |
|---|---|
| March 2012 | ByteDance founded by Zhang Yiming and Liang Rubo |
| August 2012 | Toutiao news app launched with AI recommendation |
| September 2016 | Douyin (Chinese TikTok) launched |
| September 2017 | TikTok launched internationally |
| November 2019 | CFIUS opens investigation into ByteDance |
| 2021 | Liang Rubo replaces Zhang Yiming as CEO |
| August 2023 | Doubao AI chatbot launched (invitation-only) |
| December 2021 | Volcengine public cloud launched |
| April 2024 | PAFACA signed into law requiring TikTok divestiture |
| May 2024 | Doubao Pro launched at aggressive pricing, sparking China LLM price war |
| January 2025 | US Supreme Court upholds TikTok sale-or-ban law |
| February 2025 | Wu Yonghui (ex-Google DeepMind VP) joins Seed team |
| May 2025 | Seed1.5-VL vision-language model released |
| June 2025 | Seed-Coder open-source code models released |
| August 2025 | Seed-OSS-36B open-source LLM released |
| November 2024 | Doubao reaches ~60M MAU, becomes China's top AI chatbot |
| December 2025 | TikTok US divestiture deal signed with Oracle consortium |
| January 2026 | TikTok US joint venture deal closes |
| February 2026 | Seed2.0 and Seedance 2.0 released; Doubao surpasses 100M DAU |