Google maintains official/; researchers support research/
Google describes official/ as maintained for current TensorFlow 2 APIs, while research/ contains TensorFlow 1 and 2 work supported by individual researchers. TensorFlow Model Garden collects model implementations, training drivers, configs, and checkpoints across both. community/ is a list of outside projects, and Orbit supplies training-loop utilities. A buyer should choose a directory before judging the repository because support expectations change across those four areas.
The maintained package covers three broad workloads. Vision includes classification, detection, segmentation, and video models. NLP supplies reusable layers, networks, task code, and a common training driver. Recommendation examples include ranking architectures. This breadth explains why the repository had 4,041 files and about 656,414 source lines in our checkout. It is a reference shelf for TensorFlow training work, not a single model that becomes useful after one import.
One pip route occupied 2,894 MB
Our documented pip install worked, then left 2,894 MB on disk. The root README offers tf-models-official, or a source clone added to PYTHONPATH with official/requirements.txt installed. A project that needs only an image classifier still inherits dependencies for a much wider model collection, including data, vision, NLP, optimization, and Google API packages listed in the requirements file.
Our sandbox install completed in 78 seconds, pulled 139 packages, and occupied 2,894 MB on disk. The build passed in another 6 seconds. Those are acceptable numbers on an ML workstation, but awkward for a small CI runner or a narrow service image. The repository itself was already 107.4 MB checked out, so cloning source and installing the full environment is a deliberate platform choice rather than a lightweight way to inspect one architecture.
What happened when we ran it
We cloned commit 1f08b6a and ran the project from official/ in an unprivileged Python 3.12 Debian container with 3 CPUs and 8 GB of RAM. Installation and build both succeeded. Pip-audit reported 0 known vulnerabilities in the installed Python environment. That result is useful for dependency hygiene, but it does not assess model weights, datasets, training code reached only by a specific config, or packages installed later for a particular experiment.
No test script or test target was available, so we skipped tests rather than inventing a command. Our scan found 2 CI workflow files, no Dockerfile, and no directory named tests. This does not prove the repository has no checks because test files can live beside model code and upstream automation may call private targets. It does mean the 6-second build cannot stand in for a passing suite on the measured checkout.
v2.20.0 tracks TensorFlow 2.20.0; master uses nightly packages
The stable tf-models-official package may not contain the latest master changes, according to the README. tf-models-nightly follows newer repository work. NLP users on that path are also told to install tensorflow-text-nightly. Since release 2.9, consumers can import the modeling library as tensorflow_models, but the surrounding TensorFlow and TensorFlow Text versions still need attention.
Release v2.20.0, published on February 11, 2026, targets TensorFlow 2.20.0 and points users to the matching r2.20.0 branch. Its notes also say research, tutorial, and sample models were removed from that release. An open Python 3.13 report describes installation failure with package 2.19.1 on macOS, while our Python 3.12 environment installed successfully. Pin the Python, TensorFlow, Model Garden, and TensorFlow Text versions together before debugging model code.
Training starts after installation, with data and hardware decisions
A successful import is only the entry point. Vision and NLP guides send users into model-specific configs, dataset preparation, checkpoints, training commands, and evaluation steps. Some workloads can run locally, while distributed examples target GPUs or TPUs. Dataset terms also vary: the vision guide tells users to review third-party dataset licenses, even though the provided checkpoints are covered by Apache 2.0. The package does not turn those external assets into one uniform contract.
Basic package installation required no credentials in our 78-second run. Real experiments may need access to datasets or stored checkpoints, plus enough accelerator memory for the chosen config. The repository supplies readable implementation patterns and common drivers, which is its main value. It does not provide a root Dockerfile that freezes every system dependency, and the measured environment used 2,894 MB before downloading training data or additional model artifacts.
Fresh pushes coexist with 1,073 open issues
The repository was pushed on August 28, 2026, and an issue was opened that day, so current activity is visible beyond the February release tag. GitHub reported 77,659 stars, 1,073 open issues, and 198 open pull requests. Those counts describe attention and queue size, not 1,073 confirmed defects. They also show why users should search by directory and model name instead of treating the tracker as one product backlog.
Age still shows in parts of the shelf. Open issue 13561 documents MobileNet V2 and V3 research checkpoint URLs returning HTTP 403, a concrete risk for old tutorials that depend on hosted artifacts. That does not cancel the active 2026 push history or the maintained official branch. It does support a split verdict: use official/ for serious TensorFlow training work, verify version alignment and artifacts first, and copy from research/ only when you are prepared to own that code path.