← All models

Seed GR-3

Published Publisher ByteDance

Model Positioning

Seed GR-3 is a vision-language-action (VLA) robotic manipulation model developed by ByteDance's Seed team for general-purpose robotics research. It uses a hybrid Transformer architecture that integrates vision-language and action-generation modules into an end-to-end model with approximately 4 billion parameters. (seed.bytedance.com)

Capabilities and Uses

  • Generalization and instruction understanding: In pick-and-place experiments, it handles objects and environments not seen during training, as well as instructions containing abstract concepts such as spatial relationships and size.
  • Long-horizon and dexterous manipulation: It can work with the ByteMini dual-arm mobile robot to perform tasks such as tabletop organization and hanging clothes through dual-arm coordination, including manipulation of deformable objects.
  • Rapid adaptation: Training combines robot trajectories, human trajectories collected via VR, and public vision-language data; the official report states that the model can be fine-tuned for new tasks or objects using a small number of human trajectories.
  • Experimental results: Under official test conditions, average task progress exceeded 95% for tabletop organization and was 86.7% for hanging clothes; these figures represent task progress in specific experiments and should not be treated as success rates across all scenarios. (seed.bytedance.com)

Limitations

The official source notes that the model may still make errors on unseen instructions involving new concepts and objects, and may struggle to grasp objects with unseen shapes. It may become stuck when encountering out-of-distribution situations such as objects slipping unexpectedly or environmental changes. Public API access methods and weight availability have not been verified. (seed.bytedance.com)

Release Date

2025-07-22; this date is explicitly stated in the official ByteDance Seed announcement. (seed.bytedance.com)

Pricing

As of 2026-11-18, searches of Volcengine Ark domestic pricing and BytePlus ModelArk international pricing did not verify an official RMB or US dollar price for GR-3. How inputs and outputs are billed has also not been verified, and prices for other models should not be applied. References: Volcengine Ark model pricing; BytePlus ModelArk pricing. (console.volcengine.com)

Official Sources