I’m impressed by tinygrad. Full-featured deep-learning framework with ZERO dependencies, fully hackable, and orders of magnitude fewer LoC than PyTorch.

Want to run Qwen3.8 27B? Download the quantized model from HF, git clone tinygrad, then run `DEV=METAL python3 -m tinygrad.llm --max_context 8192 -m <GGUF File>`. Read the code and experiment - takes 1 minute.

Compare that with PyTorch (great, but almost impossible to comprehend and modify, IMO) or Mojo/MAX (complex and not even fully open source, not to mention that the Modular folks aren’t particularly nice). Tinygrad is a breath of fresh air.

If you're like me and also wondered "huh, what's tinygrad up to?" apparently they now let you place $100,000 USD pre-orders on a $10M shipping container full of GPUs which they say will ship (literally it seems) in 2027.