ByteDance Seed

• Seed foundation models for general intelligence, reasoning, knowledge, coding, and multimodal understanding
• Seed2.1 agent models designed for real-world productivity and complex multi-step tasks
• Seedance video generation models supporting text, image, and multimodal video creation
• Seedream image generation models for multimodal visual creation and editing
• Seed Audio models for speech, sound effects, ambience, and scene-level audio generation
• SeedRealtime models for real-time multimodal interaction across audio, vision, and temporal signals
• Robotics research including vision-language-action models and reinforcement learning for dexterous manipulation
• AI for Science research including biomolecular structure prediction and protein design
• Research publications, technical reports, model cards, and evaluation results for Seed models
• API, playground, and application access for selected Seed models and capabilities

ByteDance Seed

ByteDance Seed is the AI research and model development team of ByteDance, focused on advancing general intelligence and developing foundation models and specialized AI systems. Its research spans large language models, multimodal intelligence, generative media, speech and audio, robotics, AI infrastructure, next-generation interfaces, and AI for science.

Foundation Models

The Seed foundation model series is designed to support general intelligence capabilities across reasoning, knowledge, multimodal understanding, coding, and agentic tasks. The model portfolio includes multiple generations, with models evolving toward more capable real-world task execution and multimodal interaction.

Seed2.1

Seed2.1 is a next-generation AI productivity model family offered in Pro and Turbo variants. It is designed for real-world productivity and complex everyday tasks, including project planning, file processing, tool use, software development, debugging, and validation.

AI Agents

Seed models are designed to support agentic workflows where an AI system can understand a task, reason through multiple steps, interact with tools, and work toward a practical result. The Seed2.1 family specifically focuses on improving the reliability of agents performing complex real-world tasks.

Multimodal Understanding

Seed’s foundation models extend beyond text to handle multiple information types. Depending on the model, capabilities include understanding text, images, video, and audio. This multimodal approach allows the models to process richer information from real-world environments and documents.

Reasoning and Knowledge

The Seed model family focuses on improving reasoning and knowledge capabilities for complex tasks. These capabilities support applications involving research, analysis, planning, problem solving, and other workflows where the model needs to interpret information before producing an answer or action.

Coding and Software Engineering

Seed2.1 includes strengthened code engineering capabilities. The model is designed to support end-to-end development workflows, including understanding requirements, generating code, debugging implementations, and validating results within real development environments.

Seedance

Seedance is ByteDance Seed’s family of video generation models. Seedance models are designed to generate video from natural-language prompts and other media inputs, with capabilities focused on motion, visual quality, prompt adherence, storytelling, and creative control.

Seedance Multimodal Video Generation

Newer Seedance generations expand video creation beyond simple text-to-video workflows. Depending on the model version, users can provide combinations of text, images, audio, and video as references or inputs for generating and editing video content.

Multi-Shot Storytelling

Seedance 1.0 introduced native multi-shot video generation. The model can create sequences containing multiple shots while maintaining greater consistency in subjects, visual style, atmosphere, and transitions between scenes.

Motion and Visual Control

Seedance models are designed to improve motion stability, physical plausibility, visual structure, and adherence to detailed prompts. The models can interpret instructions involving subjects, actions, camera movement, scene composition, and stylistic direction.

Seedream

Seedream is ByteDance Seed’s image generation model family. The models are designed for multimodal image creation and related visual workflows, with newer generations adding stronger reasoning and content creation capabilities.

Image Creation and Editing

Seed’s visual models are intended for creative image workflows involving generation, editing, visual understanding, and controlled creation. Depending on the model generation, capabilities can include working from textual instructions and reference information.

Seed Audio

Seed Audio 1.0 is an audio creation model designed to generate complete sound scenes rather than isolated audio elements. It can coordinate speech, sound effects, ambience, timing, and other audio components within a unified creative workflow.

Speech and Voice Generation

Seed Audio 1.0 can generate expressive speech based on text descriptions and authorized voice references. It is designed to maintain voice characteristics while controlling elements such as emotion, pacing, rhythm, and speaking style.

Multilingual Audio

Seed Audio 1.0 supports audio generation across more than 20 languages. The model is designed to maintain recognizable voice characteristics while adapting speech to different languages and their respective pronunciation, rhythm, pacing, and expressive requirements.

SeedRealtime

SeedRealtime is designed for natural, full-duplex interaction by jointly understanding audio, vision, and temporal information. The model focuses on real-time multimodal interaction where an AI system needs to observe, listen, and respond continuously.

Robotics Research

ByteDance Seed also conducts research in robotics. Its portfolio includes Seed GR-RL, a reinforcement learning framework focused on long-horizon dexterous manipulation, and Seed GR-3, a vision-language-action model designed for generalizable robotic control.

Long-Horizon Manipulation

Seed’s robotics research focuses on enabling robots to complete multi-step tasks involving precise physical interactions. This research combines visual understanding, language instructions, action planning, and reinforcement learning for complex manipulation scenarios.

AI for Science

Seed applies AI research to scientific problems through its AI for Science work. One area of this research is biomolecular modeling, including models designed for complex structure prediction and de novo protein design.

Protenix

Protenix is a biomolecular foundation model within ByteDance Seed’s AI for Science portfolio. It is designed for high-fidelity biomolecular complex structure prediction and supports research into protein structure and design.

Research Publications

ByteDance Seed publishes technical reports, research papers, model cards, and other documentation describing its AI systems. These publications provide information about model architecture, capabilities, training approaches, evaluations, and intended applications.

Model Portfolio

The Seed model portfolio brings together several generations and specialized model families. The portfolio includes general-purpose foundation models, generative video and image models, audio models, real-time multimodal systems, robotics models, and AI-for-science research.

Model Evaluation

Seed publishes evaluation information for selected models to describe their performance across different tasks. Depending on the model, evaluations can cover reasoning, multimodal understanding, coding, agent capabilities, video generation, image generation, audio generation, and other specialized benchmarks.

Developer Access

Selected Seed models provide API access and playground experiences. Availability depends on the individual model and the distribution platform. Developers can use supported models for integrating AI capabilities into applications and workflows.

Model Availability

Not every Seed research model is necessarily available in the same way. Some models provide public documentation or research reports, while others can be accessed through playgrounds, APIs, or ByteDance products. Availability can change as new versions are released.

AI Research Programs

ByteDance Seed also operates research programs designed to encourage long-term exploration of artificial intelligence. These initiatives include Seed Edge and programs supporting research collaboration, emerging AI directions, and future foundation-model development.

Seed Edge

Seed Edge is a research program focused on longer-term AI challenges and potential advances toward general intelligence. Its research directions include reasoning, multimodal intelligence, and other areas where substantial improvements may require longer-term investigation.

Research Areas

ByteDance Seed’s broader research portfolio covers large language models, speech, vision, world models, AI infrastructure, next-generation interfaces, robotics, generative media, and AI for science. The organization combines fundamental research with the development of models intended for practical applications.

Real-World Applications

Seed’s technologies are designed to support practical AI applications across productivity, software development, creative content generation, multimodal interaction, robotics, and scientific research. Its models have been incorporated into ByteDance products and other application environments.

Company Background

ByteDance Seed was established in 2023 as ByteDance’s dedicated AI research team. The team operates research activities across China, Singapore, and the United States and focuses on foundational AI research and the development of general-purpose and specialized AI systems.

Frequently Asked Questions

What is ByteDance Seed?

ByteDance Seed is ByteDance’s AI research and model development team. It develops foundation models and conducts research across language, vision, audio, video, robotics, AI infrastructure, and AI for science.

What models does ByteDance Seed develop?

The Seed portfolio includes Seed foundation models, Seed2.1, Seedance video models, Seedream image models, Seed Audio, SeedRealtime, robotics models such as Seed GR-3, and scientific AI systems including Protenix.

What is Seed2.1?

Seed2.1 is a foundation model family designed for real-world productivity. Its Pro and Turbo variants focus on general agents, coding, reasoning, multimodal understanding, and complex multi-step tasks.

What is Seedance?

Seedance is ByteDance Seed’s video generation model family. It supports AI-assisted video creation and, depending on the version, can work with text, images, audio, and video inputs.

What is Seedream?

Seedream is a ByteDance Seed image generation model family focused on multimodal visual creation and related image-generation capabilities.

What is Seed Audio?

Seed Audio 1.0 is an audio creation model designed to generate coordinated sound scenes containing speech, sound effects, ambience, and other audio elements.

Does ByteDance Seed provide APIs?

Selected Seed models provide API access. Availability and access conditions depend on the specific model and the platform through which it is offered.

Is ByteDance Seed free?

Seed research information, publications, and model documentation are publicly accessible. Commercial API usage and access to specific models can have separate availability and pricing conditions that vary by model.

When was ByteDance Seed established?

ByteDance Seed was established in 2023. A specific public launch date for the Seed platform itself has not been disclosed.

What research areas does ByteDance Seed cover?

Its research includes large language models, reasoning, multimodal intelligence, speech, vision, generative media, world models, robotics, AI infrastructure, next-generation interfaces, and AI for science.

Share Your Experience: