#1212th of 15 in ML frameworks and libraries
tinygrad
Small deep learning stack with tensors, autograd, and a readable compiler
Documentation ↗
(opens in a new tab)Website ↗
(opens in a new tab)Repository on GitHub ↗
(opens in a new tab)
- Stars
- 33.8k
- License
- MIT
- Last commit
- Oct 2026
- Last release
- Aug 2026
- Language
- Python
31,715 on 24 Mar 202633,753 on 13 Oct 2026
From GitHub's daily star counts.
Checked on GitHub.
Overview
tinygrad is a Python tensor library with autograd, an IR and compiler that fuses and lowers kernels, a JIT (TinyJit) with graph execution, and nn, optim and dataset modules for training. It targets OpenCL, CPU, Metal, CUDA, AMD, NV, QCOM and WebGPU, and a new backend needs about 25 low-level ops. The API follows PyTorch, but the compiler and IR are kept small enough to read and modify.
Who it is for: Engineers who want a hackable PyTorch alternative or a compiler to study
Strengths
- Whole compiler and IR are visible and hackable, unlike PyTorch
- Eager Tensor API and autograd close to PyTorch, so training loops look familiar
- Backends for OpenCL, CPU, Metal, CUDA, AMD, NV, QCOM and WebGPU
- New accelerator needs only about 25 low-level ops
Weaknesses
- Fewer functional transforms than JAX: no full vmap or pmap yet
- README says the library is not 1.0 yet
- Recommended install is from source; no PyPI instructions in the README
- Strict contribution rules: AI-looking PRs from new contributors are closed
What it needs
- GPU optional
- Models: OpenCL, CPU, Metal, CUDA, AMD
Also in ML frameworks and libraries
See all 15| Rank | Project | Score |
|---|---|---|
| 1 | transformersModel-definition library for text, vision, audio and multimodal models | 94 out of 100 |
| 2 | pytorchPython tensor library with GPU support and tape-based autograd for deep learning | 92 out of 100 |
| 3 | ultralyticsPython library and CLI for training and running YOLO vision models | 86 out of 100 |
| #4 | pytorch-image-modelsPyTorch library of image models, pretrained weights and reference training scripts | 82 out of 100 |
| #5 | tensorflowTensor and neural network framework with Python and C++ APIs | 80 out of 100 |