Back to Gap Map Model components / Robotics & embodied AI

GigaBrain

GigaAI
open weights / Overall score: 3.2

GigaBrain is GigaAI's line of vision-language-action models for robot manipulation. GigaBrain-0.7, the current release, is a 3.5B model whose three-part design combines scene understanding, prediction and action, pretrained on more than 37,000 hours of real-robot, handheld-gripper, egocentric, simulated and world-model-generated data. It ships with training configurations for AgileX, AgiBot and GigaAI's own Maker H01 robots.

GigaWorld, GigaAI's world model, and the Giga-World-Policy checkpoints are separate releases and are not part of this entry.

Openness

3 high confidence
3.0
weights
open (GigaBrain-0.7-3.5B-Base and fine-tunes on the Hub, ungated)
data
described (the 37.3k-hour pretraining mix is described, with only sample data published)
code
open (pretraining and post-training implementations with per-robot training configs)
license
Apache-2.0 (code and weights)

GigaBrain-0.7's weights and its training code are Apache-2.0. The robot data it learned from is described by source and hours, and only a small sample of it is published.

Adoption

2 medium confidence
2.0

Hugging Face downloads over the trailing 30 days for the five declared GigaBrain checkpoints, most of them for the 0.7 RoboTwin 2.0 fine-tune and base model.

Capability

4 medium confidence
4.0

GigaBrain-0.7 learns from many robot bodies and ships a two-armed policy, level with GR00T. It has no separate reasoning model or whole-body humanoid control of the kind Gemini Robotics pairs with its action model.

Verified 2026-11-12