ByteDance Seed is a research initiative focused on advancing the frontiers of artificial intelligence. They develop and release a variety of sophisticated AI models and frameworks.

Key offerings include:

• Seedance 1.5 pro: An audio-visual joint generation model for rapid video creation with integrated audio and visuals.
• Seedream 4.5: A powerful AI model.
• Seed1.5-VL: A vision-language model.
• Seed LiveInterpret 2.0: For real-time interpretation.
• Seed Realtime Voice: For real-time voice processing.

Recent research highlights and releases showcase their innovation:

• GR-RL: A reinforcement learning framework for long-horizon dexterous manipulation, enabling robots to perform multi-step, high-precision tasks.
• VeOmni: An open-source framework for training arbitrary modality models, significantly reducing development time.
• Seed-1.6-Embedding: A multimodal vectorization model for mixed modality retrieval.
• Seed3D 1.0: A foundation model for 3D generation, producing high-precision 3D models and state-of-the-art texture generation.
• Seed Diffusion Preview: An experimental diffusion language model focused on code generation with a high inference speed of 2146 tokens/s.
• GR-3: A general-purpose robot operation model supporting high generalization, long-range tasks, and flexible object manipulation with dual arms.

ByteDance Seed collaborates with industry partners, such as BYD, to accelerate AI for Science applications, like battery research. Their work spans core AI research areas including Large Language Models (LLM), Vision, Speech, Multimodal Interaction, Robotics, and Responsible AI.