smol-llama

360M parameter LLaMA trained from scratch on 6B tokens. GQA, RoPE, SwiGLU. Single H100, 22hrs, $53.

FineWeb-6B

Curated 6B token dataset from FineWeb. Pre-tokenized with custom 49K BPE vocab.

smoltorch

Autograd engine and neural networks in ~500 lines of NumPy. Educational deep learning.

NoPokeDB

Lightweight vector DB with hnswlib + SQLite. Crash recovery, 2K+ PyPI downloads.

AI Discord bot. Natural conversations, image understanding, and generation.

tinyndarray

NumPy-like ndarray in Rust with Python bindings. Learning project.

Opencode theme for VS Code

VS Code light and dark theme generated from OpenCode reference JSON.