Is GigaBrain-0.7 free to use commercially?

Yes. GigaBrain-0.7 ships under the Apache-2.0 license on both the giga-brain-0 GitHub repository and the GigaBrain-0.7-3.5B-Base model card on Hugging Face. Apache-2.0 allows commercial use, modification and redistribution as long as the license and attribution notices are kept, so a company can fine-tune the checkpoint for its own robots without a separate agreement.

How is GigaBrain-0.7 different from GigaBrain-0?

GigaBrain-0 was published as a world-model-powered vision-language-action model. GigaBrain-0.7 restructures that into a three-system architecture covering understanding, prediction and action, scales pretraining to more than 37,000 hours of heterogeneous embodied data, and adds a single-stage alignment method that trains multimodal understanding and multi-robot action generation at the same time.

Which robots has GigaBrain-0.7 been tested on?

GigaAI reports results on the Maker H01 platform and on other mainstream robot embodiments across varied environments. The technical report covers zero-shot behaviour across those bodies, language-conditioned instruction following, and post-training task success rates, which is the evidence offered for the claim that one checkpoint transfers between different robot hardware.

Where do I get the GigaBrain-0.7 weights and paper?

The GigaBrain-0.7 weights sit on Hugging Face as open-gigaai/GigaBrain-0.7-3.5B-Base, the training and inference code is in the open-gigaai/giga-brain-0 repository on GitHub, and the technical report is both a PDF inside that repository and an arXiv preprint, 2608.15875, first submitted on August 16, 2026.

GigaAI · 2026-08-25 · major

GigaBrain-0.7 — an open robot brain trained on 37,000 hours of embodied data

GigaBrain-0.7 is an open vision-language-action model for robots, released under Apache-2.0 with weights on Hugging Face. GigaAI pretrained it on more than 37,000 hours of mixed robot data and ships the code in the giga-brain-0 repo.

Hugging Face model card banner for GigaBrain-0.7-3.5B-Base, GigaAI's vision-language-action model

GigaBrain-0.7 is an Apache-2.0 vision-language-action model that turns camera views and a spoken instruction into robot movement.

Key specs

GitHub stars	2,611
Pretraining data	37,000+ hours

Quick facts

Maker	GigaAI (open-gigaai)
License	Apache-2.0
Released weights	GigaBrain-0.7-3.5B-Base
Pretraining data	37,000+ hours of embodied data
Architecture	Three-system: understanding, prediction, action
Paper	arXiv 2608.15875
Robot platform tested	Maker H01, plus mainstream embodiments

What is it?

GigaBrain-0.7 adds a three-system design that handles understanding, prediction and action in one model, and it is the first GigaBrain version to ship downloadable weights alongside the code. GigaAI published the model card as GigaBrain-0.7-3.5B-Base on Hugging Face on August 24, 2026 and pushed the matching code into the giga-brain-0 repository on August 25. The earlier GigaBrain-0 was described as a world-model-powered vision-language-action model; 0.7 is the scaled-up follow-up.

How does it work?

The three systems split the job: one reads images and language, one predicts what happens next, and one produces the actual motor commands, all inside a single network. Pretraining used more than 37,000 hours of heterogeneous embodied data — robot demonstrations gathered across home and industrial settings. The paper describes a single-stage alignment step that trains multimodal understanding and multi-robot action generation together rather than in separate passes.

Why does it matter?

Robot teams normally have to collect their own demonstration data before a policy will do anything useful, and that data is slow and expensive to gather. An Apache-2.0 checkpoint pretrained on 37,000 hours lets a lab start from a model that already generalizes across robot bodies and fine-tune from there. GigaAI reports gains in zero-shot behaviour across robot embodiments and in following spoken instructions, tested on the Maker H01 platform among others.

Who is it for?

robotics researchers and embodied-AI engineers

Frequently asked questions

Is GigaBrain-0.7 free to use commercially?: Yes. GigaBrain-0.7 ships under the Apache-2.0 license on both the giga-brain-0 GitHub repository and the GigaBrain-0.7-3.5B-Base model card on Hugging Face. Apache-2.0 allows commercial use, modification and redistribution as long as the license and attribution notices are kept, so a company can fine-tune the checkpoint for its own robots without a separate agreement.
How is GigaBrain-0.7 different from GigaBrain-0?: GigaBrain-0 was published as a world-model-powered vision-language-action model. GigaBrain-0.7 restructures that into a three-system architecture covering understanding, prediction and action, scales pretraining to more than 37,000 hours of heterogeneous embodied data, and adds a single-stage alignment method that trains multimodal understanding and multi-robot action generation at the same time.
Which robots has GigaBrain-0.7 been tested on?: GigaAI reports results on the Maker H01 platform and on other mainstream robot embodiments across varied environments. The technical report covers zero-shot behaviour across those bodies, language-conditioned instruction following, and post-training task success rates, which is the evidence offered for the claim that one checkpoint transfers between different robot hardware.
Where do I get the GigaBrain-0.7 weights and paper?: The GigaBrain-0.7 weights sit on Hugging Face as open-gigaai/GigaBrain-0.7-3.5B-Base, the training and inference code is in the open-gigaai/giga-brain-0 repository on GitHub, and the technical report is both a PDF inside that repository and an arXiv preprint, 2608.15875, first submitted on August 16, 2026.

Try it

Model: open-gigaai/GigaBrain-0.7-3.5B-Base on Hugging Face