The essence of text diffusion in ~150 lines of pure Python. Inspired by Karpathy's MicroGPT.

"Everything else is just for efficiency." — Karpathy's microgpt benchmarked across scalar autograd, NumPy, and PyTorch (RTX 5080)

gpt inference optimization game

Composable tiny intelligence from microgpt. Pure Python. Zero dependencies.

A compact Transformer model for character-level generation with Mixture-of-Experts, Rotary Attention, and F-gram contextual augmentation.

Fastest microgpt in C, C++, Rust, Zig, Odin, and Go

Pure Python Grassmann Flow language model (~130 lines, no dependencies) — "Attention Is Not What You Need" (arXiv:2512.19428) in microGPT style

MiniGPT-16MB: The Ultra-Light Storyteller, driven by a high-performance Language Model (LLM) designed to run entirely in the browser. At under 16MB, this model is optimized for low-end hardware, mobile devices, and offline use.