AI Wiki

ByteDance Seed

AI CompaniesChinese AIResearch Organizations
17 min read
Updated
Suggest editHistory
RawGraph

Last reviewed

Sources

41 citations

Review status

Source-backed

Revision

v4 · 3,347 words

ByteDance Seed
字节跳动种子团队
![ByteDance Seed logo](/wiki/file_bytedance_seed_logo1_jpg)
TypeResearch division
IndustryArtificial intelligence
FoundedEarly 2023
HeadquartersBeijing, China
Key peopleWu Yonghui (Head of Foundational Research)
Zhu Wenjia (Head of Applications)
Li Hang (Head of Research, AI for Science, Robotics, Responsible AI)
Xiang Liang (Applied Machine Learning Lead)
ParentByteDance
OwnerByteDance
ProductsDoubao-Seed-2.0 (Pro, Lite, Mini, Code)
Seed 1.5-VL
Seed 1.6
Seed-Thinking-v1.5
Seedance 2.0
Seedance 1.5 Pro
Seedream 4.0
SeedEdit 3.0
Seed-Prover 1.5
Seed-OSS-36B
Seed Diffusion
Seed Music
BAGEL
Depth Anything 3
UI-TARS-2
Seed-X-7B
Websiteseed.bytedance.com

ByteDance Seed (Chinese: 字节跳动种子团队), also known as the Seed Team, is the artificial intelligence research division of ByteDance, founded in early 2023 to build the foundation models behind Doubao, the most-used AI chatbot in China, and the Seedance video and Seedream image generators.[1][2] Seed operates research laboratories in China, Singapore, and the United States, and its models power more than 50 ByteDance products, including Doubao, Coze, and the Jimeng creative app.[2] The division describes its mission as being "dedicated to discovering new approaches to general intelligence and pushing the boundaries of AI."[2] Doubao, the consumer assistant built on Seed models, surpassed 100 million daily active users in December 2025, and its Doubao-Seed-2.0 family (released 14 February 2026) and Seedance 2.0 video model rank among the strongest systems from any Chinese AI lab.[3][9][20]

What is ByteDance Seed?

ByteDance Seed is ByteDance's foundation-model research lab, established in early 2023 as the company's response to the rapid advancement in large language models sparked by OpenAI's ChatGPT, which launched in November 2022.[1] Its remit covers the full model stack: large language models, vision-language models, video and image generation, speech, world models, theorem proving, robotics, and the training infrastructure underneath. According to the team's official site, Seed is "dedicated to discovering new approaches to general intelligence and pushing the boundaries of AI" through foundational research and development of industry-leading AI foundation models.[2]

Seed builds the model family that powers Doubao, ByteDance's flagship consumer AI assistant, which surpassed 100 million daily active users in December 2025.[3] The team's research spans large language models, computer vision, speech, multimodal interaction, AI for Science, robotics, and AI infrastructure, and its outputs reach more than 50 real-world ByteDance applications.[2]

When was ByteDance Seed founded?

ByteDance Seed was established in early 2023 as ByteDance's response to the rapid advancement in large language models sparked by ChatGPT's release in November 2022.[1] Initially known as the Doubao Seed Team, it was created to develop foundational AI models and explore new approaches to achieving artificial general intelligence.[3] In early 2024, ByteDance reorganized its AI system, spinning Seed (foundation-model research) off as an independent unit focused on building the Doubao large model.[2]

History

In October 2024, ByteDance established a joint research center with Tsinghua University's Institute for Artificial Intelligence Research (AIR) to advance industry-academia collaboration on large models.[4]

In January 2025, Reuters reported that ByteDance planned to spend more than US$12 billion on AI chips and compute infrastructure in 2025, partly via domestic suppliers, aimed at foundation-model training. While not specific to Seed, this investment underpins the company's broader model development efforts.[5]

In February 2025, ByteDance underwent a significant restructuring of the Seed division, recruiting Wu Yonghui, a former Google Fellow and Google DeepMind Vice President of Research, to lead foundational research efforts.[6] This restructuring came amid competitive pressure from DeepSeek, whose models had surpassed ByteDance's Doubao in daily active users by late January 2025.[1] In April 2025, the broader independent ByteDance AI Lab (covering AI for Science, Robotics, and Responsible AI) was fully merged into Seed, with Li Hang reporting to Wu Yonghui.[7] The same month, ByteDance integrated its robotics team into the Seed system and established Seed Robotics, oriented around general embodied intelligence.[7]

In September 2025, the team issued million-level stock options to employees to incentivize performance amid the AI talent war.[8] By December 2025, Doubao reached 100 million daily active users, achieved with the lowest marketing spend of any ByteDance product to ever cross that mark, according to internal sources cited by 36Kr and TechNode.[3]

On 14 February 2026, ByteDance and its Volcano Engine cloud arm released Doubao-Seed-2.0, a four-tier model family (Pro, Lite, Mini, Code) framed by the company as competitive with OpenAI's GPT-5.2 and Google's Gemini 3 Pro on math, coding, and reasoning, while pricing inputs at roughly an order of magnitude lower per million tokens.[9][10][11] The same launch event introduced Seedance 2.0, the audio-visual joint generation model that became Seed's most-discussed release of 2026.[12]

Organization and leadership

The Seed team operates research laboratories in China, Singapore, and the United States.[2] The division reports directly to ByteDance CEO Liang Rubo.[6] Following the February 2025 restructuring, the department is split into a foundational research arm under Wu Yonghui and an applications arm under Zhu Wenjia, both reporting to Liang.[6]

Wu has described Seed's internal structure as a three-tier system organized around timescale: Edge (long-horizon AGI exploration), Focus (current foundational-model challenges), and Base (production-ready model generations). The structure runs research and product cycles in parallel rather than in sequence.[13]

Key leadership

NameRoleNotes
Wu YonghuiHead of Foundational ResearchJoined February 2025. Spent 17 years at Google, including Google search ranking, Google Brain (2014, 2023), and Vice President of Research at Google DeepMind, where he contributed to Gemini development. PhD, computer science, University of California, Riverside (2008).[6][13]
Zhu WenjiaHead of ApplicationsNow focuses on model applications after Wu's arrival; previously led the Seed department.[6]
Li HangHead of Research, AI for Science, Robotics, Responsible AIReports to Wu Yonghui after the AI Lab merger.[7]
Xiang LiangHead of Applied Machine LearningLed development of the Doubao large language model.[1]
Huang WenhaoSenior researcherCo-founder of 01.ai; joined ByteDance in 2024.[1]

Research areas

ByteDance Seed's research spans multiple AI domains:[2]

What models does ByteDance Seed make?

ByteDance Seed develops one of the broadest model portfolios of any Chinese AI lab, spanning foundation language models (the Doubao-Seed and Seed series), vision-language models (Seed 1.5-VL), video generation (Seedance), image generation and editing (Seedream, SeedEdit), theorem proving (Seed-Prover), agents (UI-TARS), translation (Seed-X), music (Seed Music), and depth estimation (Depth Anything). The sections below summarize the flagship releases.

Foundation language models

ModelReleaseKey features
Doubao-Seed-2.0 ProFebruary 2026Frontier reasoning and agent model. Reports 98.3 on AIME 2025, 3020 Codeforces rating, 88.9 GPQA Diamond, 76.5 SWE-Bench Verified. Pricing roughly US$0.47 per million input tokens and US$2.37 per million output tokens.[10][11]
Doubao-Seed-2.0 LiteFebruary 2026General production tier balancing performance and cost.[9]
Doubao-Seed-2.0 MiniFebruary 2026High-throughput batch tier optimized for cost and latency.[9]
Doubao-Seed-2.0 CodeFebruary 2026Code-specialist tier for generation, debugging, and pull-request reviews.[9]
Seed 1.62025Multimodal model with adaptive thinking that balances task accuracy against reasoning depth.[10]
Seed 1.52025Strong performance on knowledge, code generation, and reasoning tasks.[10]
Seed-Thinking-v1.5April 2025Mixture-of-experts model achieving 86.7 on AIME 2024 and 77.3 on GPQA.[14]
Seed-OSS-36BAugust 2025Open-source LLM (Apache-2.0). 512K context, configurable thinking budget, two base variants and one instruct variant. Trained on roughly 12 trillion tokens.[15][16]
Seed DiffusionJune 2025Large-scale diffusion language model reporting 2,146 tokens per second inference, around 5.4 times faster than autoregressive baselines at comparable quality.[10]

Doubao-Seed-2.0 Pro is described internally as ByteDance's first trillion-parameter Gemini-style multimodal foundation model and the largest the team has trained since its founding.[13] It is positioned to score gold-medal levels on math olympiad benchmarks (IMO and CMO) and on the five ICPC programming competitions tested at launch.[10] On price, Seed 2.0 Pro lists at roughly one-third to one-tenth the per-token cost of comparable Western frontier models: about US$0.47 per million input tokens against US$1.75 for GPT-5.2 and US$5.00 for Claude Opus 4.5.[11]

Vision-language models

Seed 1.5-VL is a flagship vision-language foundation model composed of a 532M-parameter vision encoder and a Mixture of Experts (MoE) LLM with 20B active parameters.[17] Key achievements include:

  • State-of-the-art performance on 38 out of 60 public benchmarks at release.[17]
  • Superior performance in GUI control and gameplay scenarios compared to OpenAI CUA and Claude 3.7.[17]
  • Available via Volcano Engine API (Model ID: doubao-1-5-thinking-vision-pro-250428).[18]

The Seed1.5-VL technical report was released on arXiv (arXiv:2505.07062) in May 2025.[19]

Video generation models

Seedance is Seed's text-to-video and image-to-video product family, first released in June 2025 as Seedance 1.0. Seedance 1.5 Pro, released on 15 December 2025, was the first variant to support native audio-visual joint generation.[12]

Seedance 2.0, released on 9 February 2026, extends the family with a unified multimodal architecture that handles composition, motion, camera planning, and audio in a single generation pass:[12][20]

  • Up to 15 seconds of synchronized audio plus video output from text or image inputs.
  • Multi-shot storytelling with consistent characters and visual style across cuts.
  • Phoneme-level lip-sync in eight or more languages.
  • 1080p and selective 4K outputs with cinematic motion.

As of March 2026, Seedance 2.0 held an Elo rating of 1,269 for text-to-video and 1,351 for image-to-video on the public Artificial Analysis arena, placing it first in both categories ahead of Kling 3.0, Google Veo 3, and Runway Gen-4.5.[20] Consumer access through CapCut and Dreamina began rolling out on 24 March 2026 in Brazil, Indonesia, Malaysia, Mexico, the Philippines, Thailand, and Vietnam.[12]

The model has not been without legal controversy. The Walt Disney Company sent ByteDance a cease-and-desist letter on 13 February 2026 alleging that Seedance had been trained on Disney works without compensation. Paramount Skydance lodged a similar accusation citing Star Trek, South Park, and Dora the Explorer outputs. On 16 March 2026, U.S. Senators Marsha Blackburn and Peter Welch publicly demanded ByteDance shut Seedance down, and ByteDance paused the global rollout in mid-March following the cease-and-desist letters.[20][41]

Image generation models

ModelFeatures
Seedream 3.0Native high-resolution bilingual (Chinese, English) text-to-image model with reported 94% text rendering accuracy.[21]
Seedream 4.0Image generation with 4K resolution support and faster inference; ByteDance benchmarks above Google's Nano Banana (Gemini 2.5 Flash Image) at release.[22][23]
SeedEdit 3.0Image editing model supporting complex edits via natural language prompts with detail-preserving edits and consistent lighting.[24][25]

Specialized models

  • Seed-Prover is a lemma-style whole-proof reasoning model that refines Lean proofs iteratively using prior lemmas and self-summarization. It achieved a 78.1% success rate on past IMO problems and a silver-medal score (30 of 42) on IMO 2025, fully solving four problems and partially solving a fifth in three days.[26]
  • Seed-Prover 1.5 went further on the same competition, generating compilable Lean proofs for the first five 2025 IMO problems within 16.5 hours; the resulting 35 of 42 score corresponded to the gold-medal cutoff under the previous IMO scoring system.[27]
  • BFS-Prover is a family of theorem-proving LLMs at 7B and 32B parameters (V1-7B, V2-7B, and V2-32B), targeting verifiable formal proofs in Lean and lemma-first reasoning.[28]
  • BAGEL (BAGEL-7B-MoT) is an open-source unified multimodal model with 7B active parameters (14B total) using a Mixture-of-Transformer-Experts architecture, pre-trained on interleaved text, image, video, and web data. It outperforms Qwen2.5-VL and InternVL-2.5 on standard multimodal understanding leaderboards and reaches text-to-image quality competitive with Stable Diffusion 3 specialists. Released under Apache-2.0 in May 2025.[29][30]
  • UI-TARS-1.5 is an open-source multimodal agent for diverse tasks in virtual environments, achieving SOTA on seven conventional GUI benchmarks at release. UI-TARS-2, released on 4 September 2025, expanded the system into an all-in-one agent covering GUI, gaming, code, and tool use.[31][32]
  • Seed-X-7B is an open-source multilingual translation model series (Instruct, PPO, RM variants) released on Hugging Face under Apache-2.0.[33]
  • Seed Music is the team's AI music and audio generation model.[10]
  • Seed-LiveInterpret 2.0 is an end-to-end simultaneous interpretation model with low latency and voice cloning capabilities.[10]
  • Sa2VA combines SAM2 with LLaVA for dense grounded understanding of images and videos.[34]
  • Depth Anything 3 is the third generation of the team's monocular depth estimation models, with new variants and a streaming version released in December 2025.[35]
  • Stable-DiffCoder is a family of lightweight open-source code diffusion language models (8B base and instruct), reported as state-of-the-art among open-source models at the 8B scale.[35]

Applications and platforms

ByteDance Seed's models power over 50 real-world applications,[2] including:

  • Doubao (豆包), ByteDance's flagship AI chatbot and assistant in China, which surpassed 100 million daily active users in December 2025.[3]
  • Coze, a chatbot and application development platform for custom AI agents.
  • Jimeng (即梦), a text-to-image and video generation application launched in 2024.[36]
  • Volcano Engine (Volcengine), ByteDance's enterprise cloud platform providing API access to Seed models.
  • Feishu and Lark, ByteDance's enterprise collaboration platforms.
  • CapCut and Dreamina, video creation and editing applications, where Seedance 2.0 is being progressively integrated.[12]

Research programs

Seed Edge research program

The Seed Edge program focuses on long-horizon research toward general intelligence. It corresponds to the Edge layer in Wu Yonghui's three-tier internal structure and pursues directions less likely to ship in the next product cycle.[13][37]

Top Seed talent program

The Top Seed program recruits PhDs and interns globally to work alongside senior researchers on foundation-model and applied AI projects.[38]

Is ByteDance Seed open source?

Many of ByteDance Seed's models are open source. The team maintains an active GitHub organization and a Hugging Face presence, with most models released under the permissive Apache-2.0 license. As of early 2026, the GitHub organization had around 2,800 followers and dozens of public repositories spanning multimodal models, agents, depth estimation, theorem proving, and infrastructure.[28][30] Flagship products such as the Doubao-Seed-2.0 chat models, Seedance, and Seedream remain proprietary and are served through the Volcano Engine API, while research models like Seed-OSS-36B, BAGEL, UI-TARS, Seed-X-7B, and Depth Anything ship with open weights. Notable open-source projects include:

ProjectStars (early 2026)Focus
BAGEL~5,900Unified multimodal model (7B active, 14B total)
Depth Anything 3~5,200Monocular depth estimation
VeOmni~1,900Distributed training recipes for multimodal models
Seed1.5-VL~1,600Vision-language foundation model
Triton-distributed~1,400Distributed compiler for parallel systems
Seed-OSS-36Bactively usedLong-context open-source LLM with 512K context
Seed-Proverresearch-focusedAutomated theorem proving in Lean
Seed-X-7Bresearch-focusedMultilingual translation
Stable-DiffCoderresearch-focusedCode diffusion LLMs
UI-TARS / UI-TARS-2~27,000 (entire stack)Multimodal GUI and agent stack
Trae AgentactiveLLM-based agent for software engineering tasks[39]

At ICML 2025, the Seed team had 25 papers accepted, three of them as Spotlights, covering LLM inference optimization, speech generation, image generation, video generation and world models, and AI for Science.[40]

How does ByteDance Seed compare to other AI labs?

ByteDance Seed operates in a contested AI landscape, facing competition from:

The February 2025 restructuring was triggered in part by competitive pressure from DeepSeek, whose reasoning models gained significant share in China and briefly displaced Doubao as the most popular Chinese consumer chatbot.[1] ByteDance CEO Liang Rubo acknowledged that the company had been slow to follow up on technical directions like chain-of-thought reasoning after OpenAI's o1 model release.[1] By the end of 2025, Doubao had reclaimed the top spot in China by daily active users, crossing 100 million while keeping marketing costs low compared to historical ByteDance launches.[3]

The Doubao-Seed-2.0 launch in February 2026 marked the team's most direct positioning against frontier Western models, with Pro pricing one order of magnitude below comparable GPT-5.2 and Gemini 3 Pro tiers and roughly comparable scores on math, code, and reasoning benchmarks reported by ByteDance.[10][11] On the video side, Seedance 2.0 topped the Artificial Analysis text-to-video and image-to-video arenas at launch, the first Chinese model to lead both, before Alibaba's HappyHorse-1.0 overtook it in April 2026.[20]

Collaborations

ByteDance Seed actively collaborates with academic and industry partners. The joint research center with Tsinghua University's AIR focuses on advancing large model technologies through shared resources and expertise.[4] The division also has partnerships with universities in China, Singapore, and the United States as part of its talent development and research initiatives.[2][38]

See also

References

  1. South China Morning Post, "ByteDance restructures AI division, hiring new expert from Google amid DeepSeek pressure", 2025. https://www.scmp.com/tech/big-tech/article/3299731/bytedance-restructures-ai-division-hiring-new-expert-google-amid-deepseek-pressure ↩
  2. ByteDance Seed Team, official site. https://seed.bytedance.com/en/ ↩
  3. TechNode, "ByteDance's Doubao reaches 100M DAU with minimal marketing spend", 25 December 2025. https://technode.com/2025/12/25/bytedances-doubao-reaches-100m-dau-with-minimal-marketing-spend/ ↩
  4. Tsinghua University Institute for AI Industry Research (AIR) and ByteDance joint research center announcement, October 2024. ↩
  5. Reuters, "ByteDance plans to spend over $12 billion on AI chips in 2025", January 2025. ↩
  6. Pandaily, "ByteDance Adjusts AI Department Seed, Yonghui Wu Becomes New Head", February 2025. https://pandaily.com/bytedance-adjusts-ai-department-seed-yonghui-wu-becomes-new-head/ ↩
  7. AIBase, "ByteDance Restructures AI: ByteDance AI Lab Merges into Seed AI", April 2025. https://www.aibase.com/news/17204 ↩
  8. AIBase / 36Kr reporting on Seed stock option grants, September 2025. ↩
  9. Caixin Global, "ByteDance Unveils Doubao 2.0 AI Model to Tackle Complex Tasks", 15 February 2026. https://www.caixinglobal.com/2026-02-15/bytedance-unveils-doubao-20-ai-model-to-tackle-complex-tasks-102414865.html ↩
  10. TechNode, "ByteDance Releases Doubao-Seed-2.0, Positions Pro Model Against GPT 5.2 and Gemini 3 Pro", 14 February 2026. https://technode.com/2026/02/14/bytedance-releases-doubao-seed-2-0-positions-pro-model-against-gpt-5-2-and-gemini-3-pro/ ↩
  11. TokenMix, "Doubao Seed 2.0 Pro Review: ByteDance's $0.47 Frontier Model (2026)". https://tokenmix.ai/blog/doubao-seed-2-0-pro-review-2026 ↩
  12. Wikipedia, "Seedance 2.0". https://en.wikipedia.org/wiki/Seedance_2.0 ↩
  13. The China Academy, "Who Is Wu Yonghui? The Man Behind ByteDance's Seedance 2.0 Breakout", February 2026. https://thechinaacademy.org/meet-the-man-behind-seedance-2-0/ ↩
  14. ByteDance Seed Team, "Seed-Thinking-v1.5: Advancing Superb Reasoning Models with Reinforcement Learning", arXiv:2504.13914. https://arxiv.org/abs/2504.13914 ↩
  15. Hugging Face, "ByteDance-Seed/Seed-OSS-36B-Instruct". https://huggingface.co/ByteDance-Seed/Seed-OSS-36B-Instruct ↩
  16. ByteDance Seed Team, "Seed-OSS Open-Source Models Release", August 2025. https://seed.bytedance.com/en/blog/seed-oss-open-source-models-release ↩
  17. GitHub, "ByteDance-Seed/Seed1.5-VL". https://github.com/ByteDance-Seed/Seed1.5-VL ↩
  18. Volcano Engine model documentation for doubao-1-5-thinking-vision-pro-250428. ↩
  19. arXiv, "Seed1.5-VL Technical Report", arXiv:2505.07062. ↩
  20. Wikipedia, "Seedance 2.0" (controversy and rankings sections). https://en.wikipedia.org/wiki/Seedance_2.0 ↩
  21. ByteDance Seed Team, "Seedream 3.0" technical post. ↩
  22. ByteDance Seed Team, "Seedream 4.0" model card on Volcano Engine. ↩
  23. RecodeChinaAI Substack, "ByteDance's Gemini 3.0 Moment: Meet Seedance 2.0 and Doubao 2.0", February 2026. https://recodechinaai.substack.com/p/bytedances-gemini-30-moment-meet ↩
  24. ByteDance Seed Team, "SeedEdit 3.0" technical post. ↩
  25. ByteDance Seed Team blog index. https://seed.bytedance.com/en/research ↩
  26. ByteDance Seed Team, "ByteDance Seed-Prover Achieves Silver Medal Score in IMO 2025". https://seed.bytedance.com/en/blog/bytedance-seed-prover-achieves-silver-medal-score-in-imo-2025 ↩
  27. ByteDance Seed Team, "Seed-Prover 1.5: Advanced Mathematical Reasoning through a Novel Agentic Architecture". https://seed.bytedance.com/en/blog/seed-prover-1-5-advanced-mathematical-reasoning-through-a-novel-agentic-architecture ↩
  28. GitHub, "ByteDance-Seed/Seed-Prover". https://github.com/ByteDance-Seed/Seed-Prover ↩
  29. ByteDance Seed Team, "BAGEL: The Open-Source Unified Multimodal Model", May 2025. https://seed.bytedance.com/en/blog/seed-research-bagel-the-open-source-unified-multimodal-model-an-all-in-one-model ↩
  30. GitHub, "ByteDance-Seed organization". https://github.com/ByteDance-Seed ↩
  31. ByteDance Seed Team, "ByteDance Seed Agent Model UI-TARS-1.5 Open Source". https://seed.bytedance.com/en/blog/bytedance-seed-agent-model-ui-tars-1-5-open-source-achieving-sota-performance-in-various-benchmarks ↩
  32. GitHub, "bytedance/UI-TARS" and "bytedance/UI-TARS-desktop". https://github.com/bytedance/UI-TARS ↩
  33. GitHub, "ByteDance-Seed/Seed-X-7B". https://github.com/ByteDance-Seed/Seed-X-7B ↩
  34. ByteDance Seed Team, "Sa2VA" technical post. ↩
  35. GitHub, "ByteDance-Seed/Depth-Anything-3" and "ByteDance-Seed/Stable-DiffCoder". ↩
  36. ByteDance, Jimeng app launch announcement, 2024. ↩
  37. ByteDance Seed Team, "Seed Edge Research Program". ↩
  38. ByteDance Seed Team, "Top Seed Talent Program". https://seed.bytedance.com/en/topseed ↩
  39. ByteDance Seed Team, "Trae Agent" technical post. ↩
  40. ByteDance Seed Team, "Meet Seed at ICML 2025: 25 Papers Accepted". https://seed.bytedance.com/en/blog/meet-seed-at-icml-2025-25-papers-accepted ↩
  41. AIBase, "Seedance 2.0 Launches Globally, Tops the Artificial Analysis Video Ranking List", 2026. https://www.aibase.com/news/26464 ↩

Improve this article

Add missing citations, update stale details, or suggest a clearer explanation.

Suggest edit