NoTorch brings neural nets to pure C

The real story here is not “PyTorch is dead,” it’s that a surprisingly capable training stack can fit into a tiny C codebase when you narrow the problem enough. That makes NoTorch interesting for CPU-first experimentation, embedded-ish workflows, and people who want to understand the full training loop end to end.

• –The project bundles autograd, optimizers, quantized linear layers, tokenization, checkpointing, and transformer building blocks in a very small footprint.
• –The nanoGPT port is the strongest proof point: it trains a 10.2M-parameter model on Dracula with coherent-ish generation, which is enough to validate the stack, not enough to claim broad replacement.
• –BitNet b1.58 support is the standout technical angle, especially because it includes both forward and backward paths plus a BLAS fast path.
• –The most credible positioning is “small-model, CPU-friendly infrastructure,” not a universal deep learning platform for large-scale training.
• –The project is open source and opinionated, which will appeal to systems-minded ML engineers more than teams looking for ecosystem breadth.