CtrlK
| ByteDance Seed |
|---|
| 字节跳动种子团队 |
 |
| Type |
| Industry |
| Founded |
| ** |
| Headquarters |
| Key people |
| Zhu Wenjia (Head of Applications) |
| Li Hang (Head of Research, AI Lab Director) |
| Xiang Liang (Applied Machine Learning Lead) |
| Parent |
| Owner |
| Products |
| Seed 1.6 |
| Seed-Thinking-v1.5 |
| Seedance 1.0 |
| Seedream 3.0/4.0 |
| SeedEdit 3.0 |
| Seed-Prover |
| BFS-Prover |
| Seed-OSS-36B |
| Seed Diffusion |
| Seed Music |
| BAGEL |
| UI-TARS-1.5 |
| | 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]
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 September 2025, the team issued million-level stock options to employees to incentivize performance amid the AI talent war.[7]
The Seed team operates research laboratories in China, Singapore, the United States, and other locations globally.[2] The division reports directly to ByteDance CEO Liang Rubo.[6]
Wu Yonghui - Head of Foundational Research (joined February 2025). Wu spent 17 years at Google, including roles at Google Brain (2014-2023) and as Vice President of Research at Google DeepMind. He holds a PhD in computer science from the University of California, Riverside.[6][8]
Zhu Wenjia - Head of Applications. Following the 2025 restructuring, Zhu focuses on model application work while remaining part of the Seed department.[9]
Li Hang - Head of Research and Director of AI Lab[1]
Xiang Liang - Head of Applied Machine Learning, led development of Doubao large language model[1]
Huang Wenhao - Co-founder of 01.ai, joined ByteDance in 2024[1]
ByteDance Seed's research spans multiple AI domains:[2]
Large language models (LLMs)
Computer vision and vision-language models
Speech recognition and synthesis
Multimodal interaction and world models
AI for Science
Robotics
Responsible AI
AI infrastructure
Automatic theorem proving
| Model Name | Release Date | Key Features |
|---|---|---|
| Seed 1.5 | 2025 | Superior performance in knowledge, code generation, and reasoning tasks[10] |
| Seed 1.6 | 2025 | Integrates multimodal capabilities with "adaptive thinking" balancing model performance with reasoning capability[10] |
| Seed-Thinking-v1.5 | April 2025 | Mixture-of-Experts model achieving 86.7 on AIME 2024 and 77.3 on GPQA benchmarks[11] |
| Seed-OSS-36B | August 2025 | Open-source LLM (Apache-2.0 license) designed for long-context reasoning and multilingual tasks, three variants released[12][13] |
| Seed Diffusion | June 2025 | Large-scale diffusion language model achieving inference speed of 2,146 tokens/s, 5.4x improvement over autoregressive models[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.[14] Key achievements include:
State-of-the-art performance on 38 out of 60 public benchmarks[14]
Superior performance in GUI control and gameplay compared to OpenAI CUA and Claude 3.7[14]
Available via Volcano Engine API (Model ID: doubao-1-5-thinking-vision-pro-250428)[15]
Seedance 1.0 is a multi-task video generation model capable of:[16]
Creating 1080p videos with smooth motion and cinematic aesthetics
Native multi-shot storytelling with consistent visual style across transitions
Text-to-video and image-to-video generation
Support for diverse stylistic expressions from photorealism to animation
| Model Name | Features |
|---|---|
| Seedream 3.0 | Native high-resolution bilingual (Chinese/English) text-to-image model with 94% text rendering accuracy[17] |
| Seedream 4.0 | Advanced image generation with 4K resolution support, faster inference speed, reported superiority over Google's Nano Banana (Gemini 2.5 Flash Image)[18][19] |
| SeedEdit 3.0 | Image editing model supporting complex visual manipulations through natural language prompts with detail-preserving edits and consistent lighting[20][21] |
Seed-Prover - Lemma-style whole-proof reasoning model for mathematical theorem proving, achieving 78.1% success on formalized IMO problems, solved 5 out of 6 problems in the 2025 International Mathematical Olympiad (IMO) Grand Challenge[10][22]
BFS-Prover - Theorem proving LLMs with open models at 7B and 32B parameters (V1-7B, V2-7B/32B), targeting verifiable formal proofs in Lean and lemma-first reasoning strategies[23]
Seed-LiveInterpret 2.0 - End-to-end simultaneous interpretation model with ultra-low latency and voice cloning capabilities[10]
BAGEL - Open-source foundational model supporting multimodal understanding and generation of texts, images, and videos[2]
UI-TARS-1.5 - Open-source multimodal agent for diverse tasks in virtual environments[2]
Seed Music - AI music and audio generation model for generative audio capabilities[10]
Sa2VA - Model marrying SAM2 with LLaVA for dense grounded understanding of images and videos[24]
ByteDance Seed's models power over 50 real-world applications,[25] including:
Doubao (豆包) - ByteDance's flagship AI chatbot and assistant platform in China
Coze - AI chatbot and application development platform for creating custom AI agents
Jimeng (即梦) - Text-to-image and video generation application launched in 2024[26]
Volcano Engine (Volcengine) - ByteDance's enterprise cloud platform providing API access to Seed models
Feishu/Lark - Enterprise collaboration platforms
CapCut - Video editing application
The Seed Edge Research Program focuses on exploring the frontiers of intelligence and tackling long-term research challenges in AI toward establishing general intelligence.[27]
An initiative to attract and develop top AI research talent globally, recruiting top PhDs and interns from universities.[28]
ByteDance Seed maintains an active presence on GitHub with 36 repositories as of 2025[25] and Hugging Face with multiple public model releases.[23] Notable open-source projects include:
Seed1.5-VL vision-language model
Seed-OSS-36B large language models
BFS-Prover theorem proving models
Trae Agent - LLM-based agent for software engineering tasks[29]
Various research benchmarks and evaluation tools
ByteDance Seed operates in a highly competitive AI landscape, facing competition from:
Domestic Chinese competitors: DeepSeek, Baidu, Alibaba, Tencent
International competitors: OpenAI, Google DeepMind, Anthropic
The division's restructuring in February 2025 was partly attributed to competitive pressure from DeepSeek, whose reasoning models had gained significant market share in China.[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]
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 has also established partnerships with various institutions globally as part of its talent development and research initiatives.
ByteDance
Doubao
Large language models
Artificial intelligence in China
Video generation
Text-to-image model
Automatic theorem proving
Multimodal learning
ArXiv