tinygrad surges in GitHub stars

tinygrad is what happens when you optimize for explainability first and scale second: it is one of the best living examples of how an ML runtime actually works. The trade-off is that the same minimalism that makes it elegant also means you are signing up for alpha-grade rough edges.

• –It spans tensors, autograd, compiler/IR, JIT, nn, optim, and datasets, so it is a full stack rather than a toy autograd engine.
• –Backend support across CUDA, AMD, Metal, OpenCL, WebGPU, CPU, and Qualcomm makes portability a real differentiator, especially outside NVIDIA-heavy setups.
• –The “~25 low-level ops” philosophy is the project’s moat: it keeps the backend surface small enough that new hardware support is plausibly tractable.
• –Real-world usage in openpilot plus ongoing hardware-hacker interest shows the repo has credibility beyond tutorials and benchmarks.