ByteDance Seed

ByteDance's foundation-model research team, whose robotics effort is the GR series of VLAs.

ByteDance Seed is the foundation-model research team ByteDance formed in 2023; the Doubao large language model, the image generator Seedream, and the video generator Seedance all come out of this team. Its robotics GR series began within ByteDance's research group: GR-1 (late 2023) and GR-2 (October 2024) first went through generative pretraining on large-scale video, then were fine-tuned on robot data. Starting in 2025, releases have come under the Seed name: in July, the roughly 4-billion-parameter vision-language-action model GR-3 and the bimanual mobile robot ByteMini; in September, Robix, responsible for high-level reasoning and human-robot dialogue; and in December, GR-Dexter for dexterous hands and GR-RL, which incorporates reinforcement learning. The team has also open-sourced BAGEL, a unified multimodal model. The overall approach combines video generation, vision-language models, and real-robot data to build a general-purpose robot capable of long-horizon, fine-grained tasks.

ExampleGR-3 drives ByteMini to clear a dinner table and hang clothes on a hanger, following language instructions.

Also called

字节 Seed (short Chinese form), Seed robotics team, Seed

Related

Seed GR-3 · GR-RL · Robix · GR-1 (ByteDance) · BAGEL · Vision-Language-Action Model

Sources

ByteDance Seed 官网 (Chinese)
ByteDance - Wikipedia

As of

2026-08