SHADOW-50M: 44M param 19.8MB offline LLM
The author trained SHADOW-50M from scratch (44M parameters on 45B tokens). The model is 19.8 MB, runs ≈1,900 tok/s on laptop CPU, uses ternary weights and a large token fingerprint vocabulary, operates fully offline, and includes built‑in calculator and fast disk indexing; released as a proof of concept.