ByteDance SeedChina2025

GR-3

An AI model: a vision-language-action model (VLA) takes camera images and an instruction and turns them into movements for a robot. This page describes the software, not a robot.

Robotsnot stated

Key figures as stated by the developer

5 of 6 stated · 1 gap

Model type ⁠B vision-language-action
Controls ⁠A arms and hands whole body
Inputs ⁠A camera language robot state
Access not stated
Parameters ⁠A 4,000,000,000
Base model ⁠A Qwen2.5-VL-3B-Instruct

How to read the figures

14kgA30.9lbA: The letter in the grey square after a figure points to the source the figure came from. Robots with several sources carry several letters.
74.5lb: Imperial figures are our own conversion unless the manufacturer states an imperial figure itself. The source letter always belongs to the metric figure — a conversion has no source of its own.

The four states

Versions and notes in the key figures (5)

Model type

Developer wording: "Seed GR-3 is a new vision-language-action (VLA) model provided by ByteDance Seed.".

Controls

"It outputs an action chunk to control a bi-manual mobile robot in an end-to-end manner." Blog post: "whole-body control (WBC) based on mobile base movement".

Inputs

Technical report: "GR-3 takes as inputs a natural language instruction, observations from the environment, and the robot state." It "processes the observation images from multiple camera views and the language instruction".

Parameters

"In total, GR-3 contains 4B parameters."

Base model

The technical report names the pre-trained vision-language model of GR-3 as Qwen2.5-VL-3B-Instruct.

Full specifications

7 of 12 fields stated

The letter points to the source at the foot of the page. Fields without a figure are dashed: the developer does not state them — that is not the same as zero, and not the same as no.
Field	Value	Source and caveats
Input and output
Model type	`vision-language-action`	B Developer wording: "Seed GR-3 is a new vision-language-action (VLA) model provided by ByteDance Seed.".
Inputs	`camera` `language` `robot state`	A Technical report: "GR-3 takes as inputs a natural language instruction, observations from the environment, and the robot state." It "processes the observation images from multiple camera views and the language instruction".
Controls	`arms and hands` `whole body`	A "It outputs an action chunk to control a bi-manual mobile robot in an end-to-end manner." Blog post: "whole-body control (WBC) based on mobile base movement".
Degrees of freedom controlled	19	A For the ByteMini robot: "We use the predicted action chunk to control 19 DoFs of the robot (excluding the 3 DoFs from the lift mechanism and head) for policy rollout."
Size and speed
Parameters	4,000,000,000	A "In total, GR-3 contains 4B parameters."
Control rate	not stated
Where it runs and what it is built on
Runs where	not stated
Base model	Qwen2.5-VL-3B-Instruct	A The technical report names the pre-trained vision-language model of GR-3 as Qwen2.5-VL-3B-Instruct.
Openness
Access	not stated
License	not stated
The developer’s own claims
Learns new tasks	needs retraining	B Developer claim: "GR-3 can be efficiently fine-tuned with minimal human trajectory data, enabling rapid and cost-effective adaptation to novel tasks and objects." "With only 10 human trajectories collected with VR devices for each object".
Works across robots	not stated	Note All reported use is on the ByteMini robot of ByteDance Seed: "ByteMini, powered by GR-3 that serves as the brain of the robot". ⁠B Yes only if the developer itself says the model runs on different robot models, for example robots from two manufacturers or two of one manufacturer’s models. Several units of the same robot do not count.

Notes

ByteDance Seed describes GR-3 as a new vision-language-action model.
ByteDance Seed published the GR-3 blog post on 22 July 2025.
Country and city from the Seed website, whose content belongs to Beijing Chuntian Zhiyun Technology Co., Ltd.; the page names no headquarters for the Seed team.

Sources

A Manufacturer product page. arxiv.org/html/2507.15493v1 · Retrieved 2026-10-13
B seed.bytedance.com/en/blog/seed-research-gr-3-released-a-generalist-robot-model-for-generalization-long-horizon-tasks-and-bi-manual-deformable-object-manipulation · Retrieved 2026-10-13

One line per source, not per figure: a source has one retrieval date. The letter beside a figure points to the line it came from.