tinygrad
tinygrad.org
What it is
tinygrad is a deep learning framework for neural network training, bringing together a tensor library, compiler, JIT, and training tools. It supports forward and backward automatic differentiation, and tensor operations run lazily when a tensor is realized. Neural network classes, optimizers, and state loading and saving are included. TinyJit captures and replays kernels in a decorated function, and tensors can be distributed across multiple GPUs. Listed accelerators include OpenCL, CPU, Metal, CUDA, AMD, NV, Qualcomm, and WebGPU. The project describes use in openpilot to run a driving model on a Snapdragon 845 GPU. It supports Python and local training, with multiple deployment targets and safetensors or PyTorch weights through model-specific loaders. Installation instructions cover pip, though the project recommends installing from source. tinygrad is free under the MIT license. Its API is described as stable for a while, though the project is not yet version 1.0.
Spec sheet
- Purpose
- tinygrad is an end-to-end deep learning stack with a tensor library, compiler, JIT, and tools for training.Read 7 Oct 2026 · github.com
- Intended users
- The project describes tinygrad as a framework for deep learning and neural network training.Read 7 Oct 2026 · tinygrad.org
- License
- The GitHub repository identifies the project license as MIT.Read 7 Oct 2026 · github.com
- Accelerators
- Listed accelerators include OpenCL, CPU, Metal, CUDA, AMD, NV, Qualcomm, and WebGPU.Read 7 Oct 2026 · github.com
- Autodiff
- The project supports forward and backward passes with autodiff.Read 7 Oct 2026 · tinygrad.org
- Community support
- The project directs development discussion to GitHub and Discord.Read 7 Oct 2026 · tinygrad.org
- Installation
- The project recommends installing from source and also documents installation with pip.Read 7 Oct 2026 · github.com
- Integration
- The project says tinygrad is used in openpilot to run its driving model on a Snapdragon 845 GPU.Read 7 Oct 2026 · tinygrad.org
- JIT
- tinygrad provides TinyJit to capture and replay kernels in a decorated function.Read 7 Oct 2026 · github.com
- Lazy execution
- Tensor operations are lazy and run when the tensor is realized.Read 7 Oct 2026 · docs.tinygrad.org
- Maturity
- The documentation says tinygrad is not yet version 1.0, while noting its API has been stable for a while.Read 7 Oct 2026 · docs.tinygrad.org
- Multi GPU
- The documentation says tensors can be sharded across multiple GPUs.Read 7 Oct 2026 · docs.tinygrad.org
- Neural networks
- The library includes neural network classes, optimizers, and state load/save management.Read 7 Oct 2026 · docs.tinygrad.org
- Performance caveat
- The project FAQ says tinygrad is not yet faster than PyTorch for most use cases.Read 7 Oct 2026 · tinygrad.org
Plans and prices
tinygrad Free Open source library · installation from source or pip Read 7 Oct 2026 · github.com
Strong points and limits
Strong points
- Free plan and MIT license.
- Forward and backward autodiff are supported.
- Tensors can be sharded across multiple GPUs.
- Listed accelerators include CPU, CUDA, Metal, and WebGPU.
- Supports Python and local training.
Limits
- Not yet version 1.0.
- FAQ says it is not faster than PyTorch for most use cases.
- Project recommends installation from source.
Who it suits
tinygrad suits Python users building or training neural networks who want a free framework with multiple accelerator options. It may also suit developers interested in the project’s lazy tensor model and multi-GPU sharding.
Verdict
tinygrad combines core neural-network training tools with several accelerator options and multi-GPU sharding. Its pre-1.0 status and performance caveat are worth weighing before choosing it.
Where Specifiction places it
- Deep Learning Software#23 of 357.5