ArchestackBest of Open-Source AI on GitHub ↗3 stars (opens in a new tab)

ML frameworks and libraries

The libraries models are built and run with: tensor frameworks, model zoos and the tooling around them.

transformers leads with 92, ahead of pytorch (89) and tensorflow (80). 7 projects ranked by score. Updated from public GitHub data.

The ranking

ML frameworks and libraries: full ranking
RankProjectAdoptionFreshnessMaintenanceEasy to runAgent-readyScore
1transformersModel-definition library for text, vision, audio and multimodal models167.6k starsApache-2.0last commit Oct 20269010094–7092 out of 100
2pytorchPython tensor library with GPU support and tape-based autograd for deep learning104.8k starscustom licenselast commit Oct 20267810089–8589 out of 100
3tensorflowTensor and neural network framework with Python and C++ APIs200.9k starsApache-2.0last commit Oct 2026999770–080 out of 100
#4ultralyticsPython library and CLI for training and running YOLO vision models62.5k starsAGPL-3.0last commit Oct 20263710095–7077 out of 100
#5kerasMulti-backend deep learning library running on JAX, TensorFlow and PyTorch64.4k starsApache-2.0last commit Oct 20264610087–070 out of 100
#6scikit-learnPython library of classical machine learning algorithms built on NumPy and SciPy67.6k starsBSD-3-Clauselast commit Oct 20265610064–3069 out of 100
#7modelsReference TensorFlow 2 model implementations, plus the Orbit training-loop library77.6k starscustom licenselast commit Oct 2026679560–066 out of 100

Momentum, Verified build, Docs and Privacy are not measured yet; their weight goes to the signals shown. A dash means the signal is not scored for that kind of project. Hover a number for its rating in words.

Reviews

192 out of 100

transformers

Model-definition library for text, vision, audio and multimodal models

167.6k starsApache-2.0last commit Oct 2026

Transformers is a Python library that defines model architectures for text, vision, audio, video and multimodal models and loads pretrained checkpoints from the Hugging Face Hub. It offers a Pipeline API for inference, a PyTorch-based training API, and `transformers serve` and `transformers chat` commands. Other training frameworks and inference engines reuse its model definitions.

Strengths

  • Hundreds of architectures and 1M+ Hub checkpoints behind one unified API
  • Pipeline API covers text, audio, vision and multimodal tasks in a few lines
  • Definitions are reused by Axolotl, Unsloth, DeepSpeed, vLLM, SGLang, TGI and others
  • Model files are deliberately flat and usable independently of the library

Weaknesses

  • Not a modular toolbox of neural net building blocks, per the README
  • Training API targets Transformers' PyTorch models; generic loops should use Accelerate
  • Example scripts are only examples and may need adapting
  • Requires Python 3.10+ and PyTorch 2.6+; RAM and GPU needs are not stated
  • GPU optional
  • Docker
  • Needs Python 3.10+, PyTorch 2.6+, Hugging Face Hub
  • Models: PyTorch, Hugging Face Hub checkpoints, text, vision, audio
Documentation ↗ (opens in a new tab)Website ↗ (opens in a new tab)Repository on GitHub ↗ (opens in a new tab)
289 out of 100

pytorch

Python tensor library with GPU support and tape-based autograd for deep learning

104.8k starscustom licenselast commit Oct 2026

PyTorch is a Python package that provides NumPy-like tensor computation on CPU or GPU, plus a reverse-mode autograd system for building neural networks imperatively. It ships torch.nn, torch.multiprocessing and a DataLoader, and can be extended in Python or C/C++. GPU backends covered in the README are NVIDIA CUDA, AMD ROCm (Linux only) and Intel GPU.

Strengths

  • Dynamic graphs execute line by line, so debuggers and stack traces point at user code
  • Backends for NVIDIA CUDA, AMD ROCm and Intel GPU, plus CPU-only builds
  • Layers can be written in Python or via a C++ extension API
  • Pre-built pip/conda binaries, Docker images and Jetson wheels are available

Weaknesses

  • Building from source needs a C++20 compiler, 10 GB disk and 30-60 minutes
  • ROCm support is Linux only; Windows distributed support is described as a prototype
  • Source build with CUDA requires matching cuDNN 9.0+, NVTX and compiler versions
  • Low-level framework only; no model serving or application layer included
  • GPU optional
  • Docker
  • Needs Python 3.10+, C++20 compiler (source builds), CUDA/cuDNN, ROCm or Intel GPU stack (optional), Intel MKL
  • Models: torch.nn modules, CUDA, ROCm, Intel XPU, CPU
Documentation ↗ (opens in a new tab)Website ↗ (opens in a new tab)Repository on GitHub ↗ (opens in a new tab)
380 out of 100

tensorflow

Tensor and neural network framework with Python and C++ APIs

200.9k starsApache-2.0last commit Oct 2026

TensorFlow is a machine learning framework for building, training and deploying models, with stable Python and C++ APIs and non-guaranteed APIs for other languages. It installs via pip, with CUDA GPU support on Ubuntu and Windows and a smaller CPU-only package. Related tooling includes TensorBoard and LiteRT for Android.

Strengths

  • Stable Python and C++ APIs
  • CUDA GPU support via pip on Ubuntu and Windows; separate tensorflow-cpu package
  • Device plugins for DirectX and macOS Metal
  • Surrounding tooling: TensorBoard, LiteRT for Android, official models and tutorials

Weaknesses

  • APIs for languages other than Python and C++ have no backward compatibility guarantee
  • Built-in GPU support is CUDA only; other devices need separate plugins
  • Linux XLA build artifacts listed as TBA
  • RAM and VRAM requirements not stated in the README
  • GPU optional
  • Needs Python, CUDA (for GPU)
  • Models: LiteRT (Android)
Documentation ↗ (opens in a new tab)Website ↗ (opens in a new tab)Repository on GitHub ↗ (opens in a new tab)
#477 out of 100

ultralytics

Python library and CLI for training and running YOLO vision models

62.5k starsAGPL-3.0last commit Oct 2026

Ultralytics provides the `ultralytics` Python package and `yolo` CLI for training, validating, predicting, tracking and exporting YOLO models. Tasks include detection, instance and semantic segmentation, depth estimation, classification, pose and oriented boxes. Pretrained YOLO26 weights download on first use, and models export to formats such as ONNX.

Strengths

  • One package and CLI cover train, val, predict, track and export across seven vision tasks
  • Pretrained YOLO26 n/s/m/l/x weights with published COCO mAP and CPU/T4 latency tables
  • Supports older YOLO versions back to YOLOv3, plus export to ONNX
  • Install via pip, Conda or Docker; docs in multiple languages

Weaknesses

  • AGPL-3.0 license; commercial use needs a paid Enterprise License
  • Vision only; no LLM or language model support
  • Pretrained weights download from Ultralytics releases at first use
  • YOLO27 is announced but unavailable, with no launch date
  • GPU optional
  • Docker
  • Needs Python>=3.8, PyTorch>=1.8
  • Models: YOLO26, YOLOv3, ONNX export, TensorRT export
Documentation ↗ (opens in a new tab)Website ↗ (opens in a new tab)Repository on GitHub ↗ (opens in a new tab)
#570 out of 100

keras

Multi-backend deep learning library running on JAX, TensorFlow and PyTorch

64.4k starsApache-2.0last commit Oct 2026

Keras 3 is a high-level Python API for building and training neural networks. The backend is chosen with the KERAS_BACKEND variable or ~/.keras/keras.json. It trains on JAX, TensorFlow or PyTorch, and OpenVINO is supported for inference only. It is meant as a drop-in replacement for tf.keras and accepts tf.data pipelines and PyTorch DataLoaders as input.

Strengths

  • Same model code runs on JAX, TensorFlow or PyTorch by switching one environment variable
  • Keras models can be used inside native PyTorch Modules or JAX functions
  • Accepts tf.data.Dataset and PyTorch DataLoader inputs regardless of backend
  • Drop-in replacement for tf.keras when models are saved in the .keras format

Weaknesses

  • Backend must be set before import and cannot be changed afterwards
  • OpenVINO backend only supports model.predict(), no training
  • Only Linux and macOS are listed as compatible; Windows needs WSL2
  • Custom layers or train_step() need conversion to be backend-agnostic
  • GPU optional
  • Needs tensorflow >=2.16.1, jax >=0.4.20, or torch >=2.1.0 (one backend), openvino 2026.2.0 (inference only)
  • Models: JAX, TensorFlow, PyTorch, OpenVINO (inference only)
Website ↗ (opens in a new tab)Repository on GitHub ↗ (opens in a new tab)
#669 out of 100

scikit-learn

Python library of classical machine learning algorithms built on NumPy and SciPy

67.6k starsBSD-3-Clauselast commit Oct 2026

scikit-learn is a Python module for machine learning built on SciPy, started in 2007 as a Google Summer of Code project. It installs with pip or conda and requires Python 3.12 or newer, NumPy, SciPy, Narwhals, joblib and threadpoolctl. Plotting helpers need Matplotlib. The README does not list the algorithms; those are documented on the website.

Strengths

  • Installs with one pip or conda command; six required dependencies, all listed with minimum versions
  • BSD-3-Clause license; test suite runs with pytest sklearn
  • Volunteer community with institutional support; benchmarked with asv
  • Test randomness is controllable through the SKLEARN_SEED environment variable

Weaknesses

  • README does not mention GPU support
  • Requires Python 3.12 or newer, which excludes older environments
  • Plotting needs Matplotlib, installed separately
  • README gives no algorithm list; readers must consult the website
  • Needs Python >= 3.12, NumPy >= 1.26.0, SciPy >= 1.11.4, Narwhals >= 2.0.1, joblib >= 1.4.0, threadpoolctl >= 3.5.0
Documentation ↗ (opens in a new tab)Website ↗ (opens in a new tab)Repository on GitHub ↗ (opens in a new tab)
#766 out of 100

models

Reference TensorFlow 2 model implementations, plus the Orbit training-loop library

77.6k starscustom licenselast commit Oct 2026

Repository of TensorFlow model implementations split into official (maintained by the TensorFlow team), research (maintained by researchers, TF1 or TF2) and a curated community list. It also includes Orbit, a lightweight library for custom TF2 training loops that works with tf.distribute on CPU, GPU and TPU. Installable as tf-models-official or the nightly tf-models-nightly package.

Strengths

  • Official directory is maintained by the TensorFlow team against current TF2 APIs
  • Orbit supports custom training loops with tf.distribute on CPU, GPU and TPU
  • Published on PyPI as tf-models-official, with a daily nightly build
  • Training logs on TensorBoard.dev provided for some models

Weaknesses

  • Research directory is researcher-maintained and mixes TensorFlow 1 and 2 code
  • tf-models-official may lag master; nightly package needed for latest changes
  • Source install needs manual PYTHONPATH setup and extra requirements
  • GitHub reports license as NOASSERTION, although the README states Apache 2.0
  • GPU optional
  • Needs TensorFlow 2, tensorflow-text-nightly (NLP packages)
  • Models: TensorFlow, TensorFlow 1 (research)
Documentation ↗ (opens in a new tab)Repository on GitHub ↗ (opens in a new tab)

Written from each project's README and checked facts. Spot something wrong? Report it on GitHub (opens in a new tab).