GigaBrain-0.7 sits in a repository still named giga-brain-0

Four version numbers inside one repository name

The repository is named giga-brain-0 and its one line description calls the project a world model-powered vision-language-action model, while the headline at the top of the page announces GigaBrain-0.7 and a three-system architecture. The mismatch runs through the whole layout. The Python package directory is giga_brain_0/, the launcher is gigabrain07.py, the environment the install block creates is gigabrain07, and both mainline configs carry the gb07 prefix: gb07_pg2_pick_and_place_piper_30k.py and gb07_pg2_push_buttons_h01_30k.py.

The news list makes the lineage explicit. On 2025/11/27 came the model architecture with pre-training and post-training implementations, plus GigaBrain-0 weights. On 2026/02/02 came GigaBrain-0.1 weights, described as following the same usage as GigaBrain-0 while claiming better performance. On 2026/02/09 GigaBrain-0.1 took first place on the RoboChallenge leaderboard. On 2026/02/13 a GigaBrain-0.5M* report arrived for a variant that learns from world model-based reinforcement learning. Then the 0.7 report on 2026/08/17, 0.7 code, model and sample data on 2026/08/25, and VLM evaluation code with RoboColiseum, RoboTwin 2.0 and EBench workflows on 2026/09/02. Every box on the TODO list is checked, so the roadmap is finished rather than aspirational.

The release trail runs on Hugging Face, not on GitHub

This repository has no GitHub releases, so nothing about version 0.7 is pinned by a Git tag here. Weights arrive as four Hugging Face entries: GigaBrain-0.7-3.5B-Base as the pretrained 3.5B base model, GigaBrain-0.7-3.5B-RoboTwin2.0-Clean fine-tuned on clean RoboTwin2.0 data, GigaBrain-0.7-3.5B-EBench fine-tuned for EBench, and a GigaBrain-0.7-SampleData dataset for the training workflow.

The explanatory material is scattered the same way: a blog post at gigaai.cc/blog/gigabrain07, an arXiv entry numbered 2608.15875, and a technical report PDF at tech_report/GigaBrain-0.7.pdf inside the tree. The two fine-tuned variants are named after the benchmarks they target rather than after any data recipe, and no revision identifiers appear for them, so a rerun cannot be tied to the exact weights behind the reported figures. In practice a run is pinned on the software side by three giga packages at 1.1.0 while the checkpoint side stays whatever the hub serves at fetch time.

The dataset library version depends on your input format

Two mainline training configs are supported: gb07_pg2_pick_and_place_piper_30k.py for pick and place on PiPER, and gb07_pg2_push_buttons_h01_30k.py for button pushing on H01. Everything else, including the benchmark workflows released on 2026/09/02, carries its own installation guide and its own external dependencies. The mainline path is three pinned packages plus one requirements file, in a fresh Python 3.11.10 environment:

conda create -n gigabrain07 python=3.11.10 -y
conda activate gigabrain07

# Run from the GigaBrain-0.7 repository root.
export GIGA_BRAIN_ROOT="$PWD"

python -m pip install -r requirements-paligemma2.txt
python -m pip install "giga-datasets==1.1.0"
python -m pip install "giga-train==1.1.0"
python -m pip install "git+https://github.com/open-gigaai/[email protected]" --no-deps

export PYTHONPATH="$GIGA_BRAIN_ROOT${PYTHONPATH:+:$PYTHONPATH}"

The model package installs from a git tag with --no-deps, so pip never resolves its own dependency list. When the input data uses the LeRobot v2.1 format, the pin for giga-datasets drops from 1.1.0 to 1.0.0, and the page is explicit that this substitution is not a repository-wide benchmark requirement. So the dataset library version tracks your input format, giga-train and giga-models stay at 1.1.0, and nothing in the tree says which reader each benchmark guide expects.

requirements.txt points at a path the tree does not list

requirements.txt contains exactly one line: `-r projects/vla/giga-brain-0/requirements-paligemma2.txt`. The list of top-level repository entries has no projects directory, while requirements-paligemma2.txt does sit at the root, and the root copy is the file the install block actually installs.

pyproject.toml carries only a build-system table requiring setuptools>=61 and wheel with the setuptools.build_meta backend, and no project table, so all package metadata has to come from setup.py. That file parses requirements with a helper whose signature sets with_version=True while its own docstring says the default is False. Its line parser has a branch for lines beginning with -r that yields from a function named parse_require_file, and no definition of that name appears in the file, which ends partway through parse_requirements. MANIFEST.in and setup.cfg sit beside these two files, and neither one is described anywhere on the page.

Normalization reads Parquet directly and names no flags

Training data is expected in LeRobot format. Normalization statistics for observation.state and action come from scripts/compute_norm_stats_fast.py, which reads the LeRobot frame Parquet files directly rather than going through the dataset library. The invocation shown for it stops two lines in: the script path, a line continuation, then a bare `--` separator with no argument names following it.

So the step that has to succeed before any training run is also the one whose inputs and outputs are least specified. The benchmark guides released in September are the only places where those flags can plausibly appear, and they are separate documents with separate dependency sets. This script is not a fallback for other data layouts either: nothing here describes a reader for anything that is not LeRobot shaped, and the sample data published on the hub is what the workflow is shaped around.

Depth frames and 2D trajectories sit behind config options

The GigaBrain-0 weights released on 2025/11/27 deliberately leave out depth images and intermediate 2D manipulation trajectories, described as a more user-friendly input set. The code keeps support for both, and the route to them is the corresponding options in the configuration, valid only if your dataset carries those fields.

That makes the released checkpoint a reduced view of the architecture rather than a smaller model, and it means a dataset assembled without depth or 2D trajectories cannot exercise the code paths that read them. The later GigaBrain-0.1 weights are described as following the same usage as GigaBrain-0, so the reduced input set is the documented default across at least two weight releases. No option key is named anywhere in the repository listing, so from this repository alone there is no way to confirm which configuration entries turn each input on.

37.3k hours of data and no benchmark table

GigaBrain-0.7 is presented as an embodied foundation model with a three-system architecture unifying understanding, prediction and action, pretraining scaled past 37,000 hours, and one-stage alignment training that jointly optimizes vision-language understanding with multi-embodiment action generation. The data section gives the pretraining figure as 37.3k hours across real-robot, UMI, egocentric, simulation and world-model-generated sources.

Evaluation is on the in-house Maker H01 platform plus mainstream robot embodiments, in home and industrial scenarios, against the preceding GigaBrain-0 series and prior state-of-the-art models. That baseline is written as a raw pi entity wrapped in a subscript tag, so it does not render in plain text. The claimed gains sit in zero-shot foundation capabilities, language-conditioned instruction following and post-training task success rates, and no scores, tables or evaluation protocol appear on this page. The 2026/02/09 leaderboard result is the only concrete outcome recorded anywhere in the text, and it belongs to GigaBrain-0.1 rather than to 0.7.

Editorial conclusion

Treat giga-brain-0 as a research training stack for 0.7 rather than the 0 release its name advertises, and settle three things first: which hub checkpoint you are actually loading, which giga-datasets version your input format needs, and what the technical report says about the claimed gains, since no scores appear on this page. It fits people already training on LeRobot-format PiPER or H01 data who can read two config files. It does not fit anyone expecting a packaged model, a documented evaluation entry point, or an install that works from the repository root without dealing with requirements.txt. The last push on the default branch is 2026-09-09, and the code carries an Apache-2.0 license.

Frequently asked questions

Official sources

1. Issues
2. License: Apache-2.0
3. open-gigaai/giga-brain-0 on GitHub
4. README