Last reviewed
May 10, 2026
Sources
40 citations
Review status
Source-backed
Revision
v3 · 2,905 words
Add missing citations, update stale details, or suggest a clearer explanation.
| ByteDance Seed | |
|---|---|
| 字节跳动种子团队 | |
 | |
| Type | Research division |
| Industry | Artificial intelligence |
| Founded | Early 2023 |
| Headquarters | Beijing, China |
| Key people | Wu 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) |
| Parent | ByteDance |
| Owner | ByteDance |
| Products | Doubao-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 |
| Website | seed.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, established in early 2023 following the success of OpenAI's ChatGPT.[1] The division 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]
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 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]
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]
| Name | Role | Notes |
|---|---|---|
| Wu Yonghui | Head of Foundational Research | Joined 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 Wenjia | Head of Applications | Now focuses on model applications after Wu's arrival; previously led the Seed department.[6] |
| Li Hang | Head of Research, AI for Science, Robotics, Responsible AI | Reports to Wu Yonghui after the AI Lab merger.[7] |
| Xiang Liang | Head of Applied Machine Learning | Led development of the Doubao large language model.[1] |
| Huang Wenhao | Senior researcher | Co-founder of 01.ai; joined ByteDance in 2024.[1] |
ByteDance Seed's research spans multiple AI domains:[2]
| Model | Release | Key features |
|---|---|---|
| Doubao-Seed-2.0 Pro | February 2026 | Frontier 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 Lite | February 2026 | General production tier balancing performance and cost.[9] |
| Doubao-Seed-2.0 Mini | February 2026 | High-throughput batch tier optimized for cost and latency.[9] |
| Doubao-Seed-2.0 Code | February 2026 | Code-specialist tier for generation, debugging, and pull-request reviews.[9] |
| Seed 1.6 | 2025 | Multimodal model with adaptive thinking that balances task accuracy against reasoning depth.[10] |
| Seed 1.5 | 2025 | Strong performance on knowledge, code generation, and reasoning tasks.[10] |
| Seed-Thinking-v1.5 | April 2025 | Mixture-of-experts model achieving 86.7 on AIME 2024 and 77.3 on GPQA.[14] |
| Seed-OSS-36B | August 2025 | Open-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 Diffusion | June 2025 | Large-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]
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:
The Seed1.5-VL technical report was released on arXiv (arXiv:2505.07062) in May 2025.[19]
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]
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.[20]
| Model | Features |
|---|---|
| Seedream 3.0 | Native high-resolution bilingual (Chinese, English) text-to-image model with reported 94% text rendering accuracy.[21] |
| Seedream 4.0 | Image 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 3.0 | Image editing model supporting complex edits via natural language prompts with detail-preserving edits and consistent lighting.[24][25] |
ByteDance Seed's models power over 50 real-world applications,[2] including:
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]
The Top Seed program recruits PhDs and interns globally to work alongside senior researchers on foundation-model and applied AI projects.[38]
ByteDance Seed maintains an active GitHub organization and a Hugging Face presence, with most models released under Apache-2.0. 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] Notable open-source projects include:
| Project | Stars (early 2026) | Focus |
|---|---|---|
| BAGEL | ~5,900 | Unified multimodal model (7B active, 14B total) |
| Depth Anything 3 | ~5,200 | Monocular depth estimation |
| VeOmni | ~1,900 | Distributed training recipes for multimodal models |
| Seed1.5-VL | ~1,600 | Vision-language foundation model |
| Triton-distributed | ~1,400 | Distributed compiler for parallel systems |
| Seed-OSS-36B | actively used | Long-context open-source LLM with 512K context |
| Seed-Prover | research-focused | Automated theorem proving in Lean |
| Seed-X-7B | research-focused | Multilingual translation |
| Stable-DiffCoder | research-focused | Code diffusion LLMs |
| UI-TARS / UI-TARS-2 | ~27,000 (entire stack) | Multimodal GUI and agent stack |
| Trae Agent | active | LLM-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]
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]
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]