I trained a 44M parameter quantized LLM from scratch on 45B tokens. It ships in 19.8 MB and runs at ~1,900 tok/s on CPU. [P] | TickerVault