NOTORCH
A complete neural network framework written in pure C, with no Python or CUDA toolkit dependency, that both runs and trains models, including loading and running GGUF files directly.
NOTORCH is a from-scratch neural network framework in C that implements its own automatic differentiation, optimizers, a low-bit BitNet-style precision mode, a SwiGLU feed-forward block, a custom attention mechanism, LoRA and adapter training, an audio front end for spectral features, and alignment training methods including DPO, GRPO and distillation, compiling from a single source file in a couple of seconds with no build system dependency.
Its bundled harness loads a GGUF model file directly and generates text from a prompt or an interactive terminal chat, with support today for the Llama, Mistral, Qwen, Gemma, OLMoE and Mamba architecture families, among others, refusing by name any architecture it has not been built and tested for rather than silently forcing it through a generic forward pass. It also has a JavaScript port, notorch.js, and can be called from Python.