tinygrad
tiny corptinygrad is a deep learning framework from tiny corp, pitched between PyTorch and micrograd in size. It pairs a tensor library with autograd and nn, optim and dataset modules for training, and its own IR and compiler fuse and lower kernels that a JIT captures and replays. The whole compiler is meant to be read and changed, and it runs on CPU, OpenCL, Metal, CUDA, AMD, NVIDIA, Qualcomm and WebGPU backends.
The README recommends installing from source, so PyPI downloads undercount use. tiny corp also sells the tinybox, a deep learning computer, which is a separate product and not scored here.
Openness
5 high confidence- license
- MIT (OSI)
- source
- public
- core features withheld
- no
tinygrad is MIT licensed, and the repository publishes the whole framework, compiler and device runtimes included. No paid tier or license key gates any of it; tiny corp's business is selling hardware.
- https://pypi.org/pypi/tinygrad/json recorded 2026-11-12
PyPI metadata for tinygrad 0.14.0 declares "license_expression": "MIT".
- https://raw.githubusercontent.com/tinygrad/tinygrad/master/LICENSE recorded 2026-11-12
The LICENSE file is the MIT License text under "Copyright (c) 2024, the tiny corp": "Permission is hereby granted, free of charge, to any person obtaining a copy".
- https://raw.githubusercontent.com/tinygrad/tinygrad/master/README.md recorded 2026-11-12
The README calls tinygrad "an end-to-end deep learning stack" of tensor library, IR and compiler, JIT and nn/optim/datasets, says "the entire compiler and IR are visible and hackable", and describes no paid tier or gated feature.
Adoption
2 low confidencePyPI downloads of tinygrad. The README recommends installing from source or straight from the repository, so the package count is a floor on use rather than a measure of it.
- https://pypistats.org/api/packages/tinygrad/recent recorded 2026-11-12
"last_month":28545 downloads for tinygrad.
Capability
4 medium confidencetinygrad trains and runs models end to end on its own compiler across eight families of hardware, which puts it among the foundational frameworks. It sits a step below PyTorch, whose distributed training and ecosystem it does not match, and its own README says it has fewer functional transforms than JAX.
- https://raw.githubusercontent.com/tinygrad/tinygrad/master/README.md recorded 2026-11-12
The README lists "Tensor library with autograd", "IR and compiler that fuse and lower kernels", "JIT + graph execution" and "nn / optim / datasets for real training"; "BEAM search over kernels"; "Fewer functional transforms (no full vmap/pmap yet)"; accelerators OpenCL, CPU, METAL, CUDA, AMD, NV, QCOM and WEBGPU; and "only needs to support a total of ~25 low level ops".
Verified 2026-11-12